A Machine Learning Based Music Retrieval And .

1y ago
811.82 KB
8 Pages
Last View : 2m ago
Last Download : 6m ago
Upload by : Laura Ramon

A Machine Learning based Music Retrieval and Recommendation SystemNaziba Mostafa, Yan Wan, Unnayan Amitabh, Pascale FungHuman Language Technology CenterDepartment of Electronic and Computer EngineeringThe Hong Kong University of Science and Technology, Clear Water Bay, Hong Kongnmostafa@connect.ust.hk, ywanad@connect.ust.hk, uamitabh@connect.ust.hk, pascale@ece.ust.hkAbstractIn this paper, we present a music retrieval and recommendation system using machine learning techniques. We propose a query byhumming system for music retrieval that uses deep neural networks for note transcription and a note-based retrieval system for retrievingthe correct song from the database. We evaluate our query by humming system using the standard MIREX QBSH dataset. We alsopropose a similar artist recommendation system which recommends similar artists based on acoustic features of the artists’ music, onlinetext descriptions of the artists and social media data. We use supervised machine learning techniques over all our features and compareour recommendation results to those produced by a popular similar artist recommendation website.Keywords: query by humming, similar artist recommendation, music information retrieval, machine learning1.IntroductionFaster computational speed and increasing number of online users have resulted in a dramatic increase in musicconsumption. It is getting more and more difficult for thegeneral public, especially non-experts, to find and retrievemusic from the millions of songs available online. A lotof research is being done these days to find efficient music retrieval and recommendation methods. One music retrieval method that is gaining a lot of popularity these daysdue to its convenient usage is query by humming, whichis a content-based music retrieval method that can retrievemelodies using users’ hummings as queries. This allowsusers to find old songs that they only remember the tune ofor retrieve obscure songs heard in public places. The Music Information Retrieval (MIR) community has also beendoing a lot of work on automatic recommendation systemsranging from the content-based methods to social taggingand similarity networks (Cohen and Fan, 2000; Hong et al.,2008). One of the key research topics in this area that hasgained a lot of traction is automatic similar artist recommendation.Currently, there are several musical retrieval and similarartist recommendation apps. There are apps such as SoundHound, MusixMatch etc, that can retrieve songs using humming as a query, and websites such as All Music Guide(AMG)1 and last.fm2 that give similar artist recommendation. However, accuracy and efficiency of these music retrieval and recommendation systems still leave a lot of roomfor improvement. Therefore, we are planning to create aholistic music retrieval and recommendation system usingmachine learning techniques.The biggest challenges of a query by humming systeminclude i) queries sung by users often vary from the actual melody in pitch, tempo etc. so the melodic similaritymatching must be done at a more abstract level in order toget meaningful results, ii) background noise is often presentin users’ queries which also makes it harder to identify ody correctly and iii) efficient retrieval methods must beused that can search through a database and retrieve the correct melody in as little time as possible. Therefore, methodsused to retrieve the melody in this case need to be robust tonoise and inaccuracies in the singing or humming which isvery challenging, and, for the system to be practical, theentire system should be very fast.Therefore, we propose a supervised machine learning basedmethod for the query-by-humming system, which can learnthe common errors associated with human humming andbuild a model that is unaffected by these errors. For thistask, we have collected humming data and transcribed themin order to train a Deep Neural Network (DNN) based Hidden Markov Model (HMM) for note transcription. Thisdeep learning method allows us to learn the patterns presentin the humming data and create a model that can detect thenotes in a humming query. The proposed note transcriptionmethod is used along with a note-based retrieval methodsimilar to Yang et al. (2010) in order to retrieve a rankedlist of songs most similar to the query.One of the biggest challenges faced by the current similarartist recommendation systems is that they perform poorlyfor relatively obscure artists. Therefore we are interested inusing machine learning methods to build a recommendationsystem that can provide good similar artist recommendation even for relatively unknown artists. We propose a recommendation system that uses supervised machine learning techniques over features such as acoustics of music, themeta-data, and online texts related to the artist to find similar artists.2.Previous WorkThe main components of a QBH system consist of i) representation of the query and actual songs and ii) retrievalthe songs efficiently and accurately from the database. Asong or a query is mainly represented using frame-basedand note-based methods. The frame-based methods use arepresentation of the extracted pitch to represent the queryand the songs and then use some template-matching similarity measures such as DTW (Dynamic time warping) to1970

measure similarity between the main songs and the query(Wang et al., 2008; Dannenberg et al., 2007). The notebased methods use the pitches as features to transcribe thenotes present in a query or a song (Shih et al., 2002; Shih etal., 2003; Shifrin et al., 2002) and the notes in the query arethen matched against the notes in songs using simple stringmatching techniques (Ghias et al., 1995; Shih et al., 2002)or linear scaling based methods (Yang et al., 2010).In this paper, we focus on note-based methods since theyare more efficient (Kharat et al., 2015; Yang et al., 2010)and there still seems to be room for improvement in accuracy in this case. These methods are often based on statistical approaches. Hidden Markov model (HMM) is oneof the common methods that have been used for note transcription (Shih et al., 2002; Shih et al., 2003; Shifrin etal., 2002; Ryynänen and Klapuri, 2004). In Shih et al.(2002) and Shih et al. (2003), the note is segmented bymodelling phonemes using mel-frequency cepstral coefficients, energy measures, and the derivatives of these as features, and then the average pitch of the note is found onthe segmented segments, which is then used to representthe individual notes. However, this approach works wellwhen each note is hummed using one syllable such as daor ta and is not very effective for handling a large varietyof queries. The most effective of these statistical melodytranscription approaches is proposed in Ryynänen and Klapuri (2004), which extracts prosodic features, that are usedto train HMM-GMMs for modelling notes. Since recentstudies in speech recognition field have shown that usingDNN instead of GMMs in HMM significantly improves therecognition accuracy (Dahl et al., 2012), we propose to useDeep Neural Networks (DNN) with HMMs instead of simple HMM-GMM models as humming is similar to speech.The methodology used in our artist recommendation system differs from previous work both in the source and target. We collect tons of news from mainstream news websites and calculate the co-occurrence of Bollywood artists’names in these articles, which is a plausible profound andcomprehensive way to tell the relativeness of two singers.3.MethodologyThe overall system takes a hummed tune as an input, whichis then fed to the Query by Humming (QBH) system. TheQBH system uses the input to output a ranked list of songswith highest similarities to the query. The user can theneither manually choose the correct song from the rankedlist or use the default setting of choosing the most highlyranked song as the song to be retrieved. The retrieved songalong with its metadata is then used as an input to the similar artist recommendation system, which then outputs a listof most similar artists. An overview of the overall systemis given below in Figure 1.Figure 1: Overview of the overall music retrieval and recommendation systemThe Query by Humming system and similar artist recommendation system are described in more detail in sections4 and 5 respectively.4.Query by HummingThe Query by Humming system takes a few notes from amelody hummed or sung by the user as the query. Thenotes of the query is transcribed using our note transcription method and is then passed onto the retrieval system,which uses the transcribed query and the melody database,which refers to the entire list of pre-transcribed melodiesor songs that can be recognized by our system, to give aranked list of melodies that match the input query.On the other hand, related artists ought to influence eachother in their musical style. Therefore, we are also extracting audio-based features to find related artists. Voicefeatures such as Mel Frequency Cepstrum Coefficients(MFCCs) (Mermelstein, 1976) are wildly used in speechrecognition and audio fingerprinting (Cano et al., 2005).Features of MFCCs include spectral flatness, tone peaks,which could represent the features and categories of thesongs. In addition, musical features like loudness, pitchand brightness are also used for query of music (Wold etal., 1996). Su et al. (2013) have previously investigatedpiece-level features for determining the mood of a musicalpiece with high accuracy.Figure 2: Overview of the QBH systemThe rest of the paper is organized as follows. Section 3describes the overall methodology, Section 4 describes thequery by humming system, Section 5 describes the similarartist recommendation system, Section 6 explains the experimental setup and evaluation of the system and Section7 summarizes the content of the paper4.1.Note Transcription4.1.1. Feature ExtractionAs mentioned earlier, pitch is the most important characteristic of the melody. Currently, none of the pitch extrac-1971

tion algorithms is completely accurate. Therefore, we decided to use three of the best pitch extraction algorithmsaccording to Molina et al. (2014) as features to improveour systems accuracy. Those features include the pitchyinfft (Brossier, 2006), melodia (Salamon et al., 2014) andpyin (Mauch and Dixon, 2014) algorithms.first estimates the key of the musical piece. Then different note bigrams are defined for each key. Therefore, giventhe previous note i and the estimated key k, the note bigram probability P (n j n i, k) gives the probabilityof moving from note i to note j. Acoustic ModellingFor this task, we propose to train notes in the range of 35-85since this range generally covers all the notes used for human humming. We use 3-state HMM monophone modelsto train each of the notes and a single-state HMM to trainthe silence model.The extracted features are then used to train the models byusing the greedy layer-wise supervised training (Dahl et al.,2012) method, which takes the extracted features as inputand uses three hidden layers for training, which was foundto be the optimal number of layers for this task. The DNNis trained using the Kaldi toolkit 3 .An overview of the acoustic model is shown in Figure 3.Candidate Melody RetrievalThe final step is to retrieve the candidate melody represented by the hummed query. For this purpose, the melodycontour of the query is matched against those of all thesongs in the database. The melodies ranking the highestsimilarity scores are presented as the candidate melodies.The retrieval method used is similar to Yang et al. (2010).It mainly uses note-based linear scaling (NLS) and notebased recursive alignment (NRA). It uses the pitch and timeinformation of the note and recursive-alignment combinedwith linear scaling to match the query with the melody.However, instead of using absolute pitch values like in Yanget al. (2010), we use the note transition values to match thesimilarities.The note based linear scaling algorithm basically uses different scale factors to stretch and contract the hummingquery input. The distance between the humming and thesong is calculated by adding those between all the intervals.The smallest distance is then used. The basic principle behind the note based linear scaling method is shown usingFigures 4 and 5. The same humming query is used in bothfigures with different scaling and the main melody. As itcan be seen from the figures, the scaling of the query has ahuge impact when we calculate its distance from the mainmelody.Figure 3: Overview of the acoustic modelling of the notetranscription system4.1.3. Musicological ModellingThe musicological model controls transitions among thenote models and the rest in a manner similar to the languagemodel used in speech recognition.The transition probabilities among note HMMs are defined by note bi-grams, which were estimated from alarge database of MIDI files containing melodies similar toRyynänen and Klapuri (2004). Since key is important in determining note transitions as some note sequences are morecommon than others in a certain musical key, the model3http://kaldi.sourceforge.net/Figure 4: Principle behind NLSThe note based linear scaling distance calculates the globaldistance between humming and the main song. Note basedrecursive alignment is used for the local alignment. Linearscaling using a single value is generally not so effective because the duration of the note segments often varies greatly.Therefore, the humming query input is generally dividedinto several segments and linear scaling is used on each ofthe segment to get the optimal distance between the query1972

Figure 5: Principle behind NLSand the melody. Figure 6 shows the general principles ofNRA.glish and Hindi. They are NDTV4 , The Indian Express5 andWikipedia6 . We have downloaded 3431 articles, 2622 fromIndian Express and 809 from NDTV and Wikipedia. Thearticles are very comprehensive in the scope of the newsthey covered, including artists influences, collaborative efforts, gossip news, etc.In order to evaluate our results, we need to manually build astandard related artist set for each artist. Regretfully, thereis no acceptable gold standard online for Bollywood artistsand the information available on AMG is very limited inrecommending similar artists for lots of singers. We selected three Indian students with a strong Indian musicalbackground as the annotators, and provided them with thefull artists name list. They independently chose the similar artists for each target artist. We asked them to choosearound 10 related artists for each candidate and pick theones they all agree with in order to show a fair comparisonto the baseline, Last.fm, which shows around 10 similarartists for each target artist. Last.fm7 , a popular internetradio, and online music service, uses classification of metadata tags to find similar artists. Since its data is relativelyopen, it is one of the standards that music information retrieval work compares their results to.Figure 6: The procedure of NRAThe candidate songs are then ranked by their smallest distance with the query, with the song with the lowest distance ranked first. The retrieval method is used to generatea ranked list of top 20 songs most similar to the query.Figure 7: Spectral Mean Features with Principle Component Analysis5.2.5.Similar Artist RecommendationWe test the similar artist recommendation system on Bollywood artists mostly because Bollywood music provides uswith a comparatively small set, which is easier to annotateand evaluate.5.1.Spectral Mean Distance of Audio FeaturesBollywood music is influenced by both classical Indian music and modern western music. A particular Bollywoodartist is a composer or singer with his/her own style. Forexample, Rahat Fateh Ali Khan is known to fuse devotionalMuslim Sufi music with other styles. We propose that musical characteristics of an artist can be represented as theDataset Building4Bollywood industry is a relatively small circle with a total number of 116 artists. There are three main websitesthat introduce and discuss Bollywood artists in both .com6http://www.wikipedia.org7http://www.last.fm5

aggregated average acoustic features of all his/her songs.We can then measure the similarity between two artists using spectral distance measurements.We extracted musical (Tzanetakis and Cook, 2000), psychoacoustic (Cabrera, 1999) and speech features (Eyben etal., 2010) from Bollywood songs for each artist in a spectralvector representation. Su et al. (2013) used these featuressuccessfully in categorizing musical genres and moods.The musical features include timbre, chroma, spectral flatness; psychological features include loudness, sharpness;sound features include frequency and speech characteristics. For each artist si , the feature dimension is 865. Eachentry vi (k) is the mean value of the corresponding featurefor artist si . The distance of two artists over the acousticfeature space is calculated as follows:d(i, j) kvi vj k(1)where i, j stand for two artists, d(i, j) is the distance ofartist i and j over acoustic feature space, k · k is the L2norm, vi is the audio feature vector of artist i, dimension is865.Note that each feature in the acoustic space has been normalized by its mean and variance. Thus the closer the distance is, the more similar the styles of the songs of twoartists are.5.3.Figure 8: Degree: Link among Bollywood artistswhen moving the window throughout the whole articles.The degree for the artist is calculated as follows:Co-occurrence in the textsWe extract the co-occurrence of two artists in the contexts.For two arbitrary artists, si and sj , the co-occurrence iscomputed as follows:cT cjco(i, j) i ci cj (2)where co(i, j) is the co-occurrence score of artist i and j.ci is the number of times artist i occurs in each window, · is the L1-norm. For simplicity, we set our window size tothe length of each paragraph.5.4.Degree of related artistsnameLata MangeshkarA R RahmanKishore KumarVijay BenedictVijay 3k53.1k1151161.25k2.06kri The co-occurrence score measures the closeness of twoartists in the text. In this section, we propose a new feature called, degree. In Graph theory, the degree means thenumber of edges that are incident on the vertex. Analogousto this definition, we define the degree of an artist as thenumber of times that the other artists are ”incident on” theartist. Given any artist as the vertex, we calculate all thetimes that the other artists co-occur in the same paragraph(ck (i, j) 0 ? 1 : 0)(3)j6 i;kwhere ck (i, j) indicate artist i and j co-occur in window k.In our experiment, artists with the highest degree and lowest degree can be viewed from Table 1. We have crossreferenced the results with Last.fm play counts, which areshown in Table 1 Listeners column, for the artists and wehave found that artists with the higher degree have largerplay counts, which means they are more popular. Generally speaking, the artist with a higher degree tends to bemore influential. This feature helps us to re-rank the candidate list and balance the results with more popular andlesser-known artists.5.5.Table 1: The rank of artists sorted by their degree. Listenersare the data from Last.fmXLearning Feature WeightsWe now have three categories of features for our trainingsystem. They are co-occurrence of this artist with the target artist, degree of influence of this artist, and the spectralmean distance of the artist with target artist. We constructour training data set as follows. For the total number of 116artists, we construct tuple sets. For artist si , we have Ĉi andCi , where Ĉi means calculated candidate artists set that isrelated to si and Ci means standard candidate artists set thatis related to si . Thus, for each sj Ĉi , if sj 6 Ci , we canbuild a tuple with the flag ”false”. Otherwise, the flag is setto ”true”. Each tuple contains four elements, that includethree features and one flag. Once we get all the tuples, wesplit it into training set and testing set with the ratio 9:1. Weuse 10-fold validation for average performance. Since it isan unbalanced tuple set with negative tuples being dominant, we select negative tuples uniformly at random to the1974

DegreeHighLowcandidate setaritist nameLata MangeshkarA R RahmanKishore KumarSonu Nigam.Shubha MudgalShibani .000.750.630.800.840.630.58Co-occurrencePRF10.08 0.150.13 0.11 0.121.00 0.33 0.500.20 0.10 0.13.00000010.13 0.220.31 0.06 0.10P0.730.450.640.180.730.730.550.55Spec CoRF10.33 0.460.56 0.500.58 0.610.20 0.190.891.000.750.610.800.840.630.56Spec Co DegreePRF0.72 0.33 0.460.64 0.78 0.700.55 0.50 0.520.36 0.40 55Table 2: Results (precision, recall, F-score) in percentage for comparing three features performances over the last.fmresults. It is calculated by 10-fold validation. Each fold iterates for 100 times in logistic regression.Extracted FeaturesSimple autocorrelation based pitch extractionOnly ”PitchYinfft” algorithmPitchYinfft, Melodia and pYin algorithmsMRR0.4870.690.8071Table 3: Comparison of results with pitch obtained usingdifferent pitch-extraction algorithms as featuresNote Transcription AlgorithmHMM-GMM based acoustic modelDNN-HMM based acoustic modelMRR0.76790.80716.2.Table 4: Comparison of results with different algorithmssame size of the number of positive tuples to make a balanced set. We use logistic regression to train the model andupdate the weights with the stochastic gradient descent.6.6.1.Experiments and ResultsResults for Humming RecognitionFor training note models, we have used humming data fromthe IOACAS corpus 8 and TCS corpus 9 with additionalhumming data collected by us. We annotated the hummingdata manually using the Tony software 10 .For the evaluation of the overall query by humming system, we have used the standard set used by MIREX for thispurpose. It uses the Roger Jang corpus 11 , consisting of4431 queries and 48 ground-truth MIDI files. So, for ourexperiments we first transcribe the notes in all the groundtruth MIDI files. The queries are then each transcribed andpassed onto our retrieval system, which generates a list ofmost likely candidate melodies. The system is evaluatedusing mean Reciprocal Ranking (MRR):M RR Q 1 X(1/ranki ) Q indicates that using a combination of pitch values as features give better results. We also first created a simpleHMM-GMM based model and the results in Table 4 indicate that using Deep Neural Networks (DNN) with HMMimproves the overall retrieval rate of the system. We havecurrently trained the transcription system on a relativelysmall dataset, and we believe that using additional training data can improve our overall transcription accuracy andthe retrieval accuracy.(4)Results for Related Artist RecommendationOnce we have done the logistic regression, we can applythe trained weights to the test set. We evaluate each artist siand calculate the related artists set Ĉi . Then we compare itto the standard candidate set Ci and calculate the precision,recall and F-score.Table 5 shows parts of our results. There are five columns.The first column is the baseline of Last.fm’s results. Theother four are the results from combination of features. Section 5.4. shows that artists with higher degree contain morecorrelation links to other artists, which partly reflect theirinfluences. We show artists with the highest degree andlowest degree. From Table 5, we can see that Last.fm doesnot work very well for artists with low degree. Actually,we can not find similar artists information for artists withlow degree on the last.fm’s website. On the contrary, ourmethod compensates this shortage. We can see that ourmethod performs smoothly when dealing with both highdegree and low degree artists and performs 40% better onaverage in F-measure. Actually, the sole spectral mean distance feature has already reached a pretty good precisionand recall for some artists. Combined with co-occurrencefeatures, we can see that precision and recall increased forthe high degree artists whereas they did not decrease onthe low degree artists. However, the co-occurrence featurealone does not perform so well. This may be due to the factthat our corpus is not large enough, so we will continue tocollect data in order to improve our results in the future.(i 1)We initially used different feature sets and the MRR obtained using the different features is shown in Table 3. w.music-ir.org/mirex/wiki/97.ConclusionIn this paper, we present a music retrieval and recommendation system using machine learning techniques. Wepropose a Deep Neural Network (DNN) based note transcription method and create a complete query by hummingmusic retrieval system, which we test using the standardMIREX Query by humming data set. We show that the1975

DegreeHighLowcandidate setaritist nameLata MangeshkarA R RahmanKishore KumarSonu Nigam.Shubha MudgalShibani .000.750.630.800.840.630.58Co-occurrencePRF10.08 0.150.13 0.11 0.121.00 0.33 0.500.20 0.10 0.13.00000010.13 0.220.31 0.06 0.10P0.730.450.640.180.730.730.550.55Spec CoRF10.33 0.460.56 0.500.58 0.610.20 0.190.891.000.750.610.800.840.630.56Spec Co DegreePRF0.72 0.33 0.460.64 0.78 0.700.55 0.50 0.520.36 0.40 55Table 5: Results (precision, recall, F-score) in percentage for comparing three features performances over the last.fmresults. It is calculated by 10-fold validation. Each fold iterates for 100 times in logistic regression.QBH system overall shows encouraging results and can beimproved with additional data. We also propose a similarartist recommendation system and experiment the systemon an exhaustive list of 116 Bollywood artists and showthat the recommendations based on spectral distance, cooccurrence and degree measures give better results on average for all artists compared to popular similar artist recommendation website. We plan to collect more data in thefuture and test the system on a larger dataset.8.AcknowledgementsThis research was partially supported by Grant Number16214415 of the Hong Kong Research Grant Council andpartially supported by Bai Xian Asian Institute. We wouldlike to thank our colleagues from Human Language Technology Center of The Hong Kong University of Science andTechnology, who provided expertise that greatly assistedthe research. We thank Anik Dey for assistance with hisBollywood knowledge which improved the precision of ourwork and Ricky Chan for assistance with the acoustic modeling of note transcription system. We would also like tothank the three annotators who helped us label our dataset.9.Bibliographical ReferencesBrossier, P. M. (2006). Automatic annotation of musicalaudio for interactive applications. Ph.D. thesis, QueenMary, University of London.Cabrera, D. (1999). Psysound: A computer program forpsychoacoustical analysis. In Proceedings of the Australian Acoustical Society Conference, volume 24, pages47–54.Cano, P., Batlle, E., Kalker, T., and Haitsma, J. (2005). Areview of audio fingerprinting. Journal of VLSI signalprocessing systems for signal, image and video technology, 41(3):271–284.Cohen, W. W. and Fan, W. (2000). Web-collaborativefiltering: Recommending music by crawling the web.Computer Networks, 33(1):685–698.Dahl, G. E., Yu, D., Deng, L., and Acero, A. (2012).Context-dependent pre-trained deep neural networks forlarge-vocabulary speech recognition. Audio, Speech,and Language Processing, IEEE Transactions on,20(1):30–42.Dannenberg, R. B., Birmingham, W. P., Pardo, B., Hu, N.,Meek, C., and Tzanetakis, G. (2007). A comparativeevaluation of search techniques for query-by-hummingusing the musart testbed. Journal of the American Society for Information Science and Technology, 58(5):687–701.Eyben, F., Wllmer, M., and Schuller, B. (2010). Opensmile: the munich versatile and fast open-source audiofeature extractor. In Proceedings of the internationalconference on Multimedia, pages 1459–1462. ACM.Ghias, A., Logan, J., Chamberlin, D., and Smith, B. C.(1995). Query by humming: musical information retrieval in an audio database. In Proceedings of the thirdACM international conference on Multimedia, pages231–236. ACM.Guthrie, J. A., Guthrie, L., Wilks, Y., and Aidinejad,H. (1991). Subject-dependent co-occurrence and wordsense disambiguation. In Proceedings of the 29th annualmeeting on Association for Computational Linguistics,pages 146–152. Association for Computational Linguistics.Hong, J., Deng, H., and Yan, Q. (2008). Tag-based artistsimilarity and genre classification. In Knowledge Acquisition and Modeling Workshop, 2008. KAM Workshop2008. IEEE International Symposium on, pages 628–631. IEEE.Kharat, V., Thakare, K., and Sadafale, K. (2015). A surveyon query by singing/humming. International Journal ofComputer Applications, 111(14).Mauch, M. and Dixon, S. (2014). pyin: A fundamental frequency estimator using probabilistic threshold distributions. In Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on,pages 659–663. IEEE.McDermott, J. H. and Oxenham, A. J. (2008). Music perception, pitch, and the auditory system. Current opinionin neurobiology, 18(4):452–463.Mermelstein, P. (1976). Distance measures for spe

Department of Electronic and Computer Engineering The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong nmostafa@connect.ust.hk, ywanad@connect.ust.hk, uamitabh@connect.ust.hk, pascale@ece.ust.hk Abstract In this paper, we present a music retrieval and recommendation system

Related Documents:

Machine learning has many different faces. We are interested in these aspects of machine learning which are related to representation theory. However, machine learning has been combined with other areas of mathematics. Statistical machine learning. Topological machine learning. Computer science. Wojciech Czaja Mathematical Methods in Machine .

Machine Learning Real life problems Lecture 1: Machine Learning Problem Qinfeng (Javen) Shi 28 July 2014 Intro. to Stats. Machine Learning . Learning from the Databy Yaser Abu-Mostafa in Caltech. Machine Learningby Andrew Ng in Stanford. Machine Learning(or related courses) by Nando de Freitas in UBC (now Oxford).

Machine Learning Machine Learning B. Supervised Learning: Nonlinear Models B.5. A First Look at Bayesian and Markov Networks Lars Schmidt-Thieme Information Systems and Machine Learning Lab (ISMLL) Institute for Computer Science University of Hildesheim, Germany Lars Schmidt-Thieme, Information Systems and Machine Learning Lab (ISMLL .

Artificial Intelligence, Machine Learning, and Deep Learning (AI/ML/DL) F(x) Deep Learning Artificial Intelligence Machine Learning Artificial Intelligence Technique where computer can mimic human behavior Machine Learning Subset of AI techniques which use algorithms to enable machines to learn from data Deep Learning

Music is an important part of school life 1. Students enjoy making music and many continue to play music throughout their lives after a good music education in school 2. Music programs can bring the school community together, raise the school’s profile in the community, and boost morale Music education can have benefits to other areas of learning

a) Plain milling machine b) Universal milling machine c) Omniversal milling machine d) Vertical milling machine 2. Table type milling machine 3. Planer type milling machine 4. Special type milling machine 5.2.1 Column and knee type milling machine The column of a column and knee

Machine Learning and Econometrics This introductory lecture is based on –Kevin P. Murphy, Machine Learning A Probabilistic Perspective, The MIT Press, 2017. –Darren Cook, Practical Machine Learning with H2O, O'Reilly Media, Inc., 2017. –Scott Burger, Introduction to Machine Learning

2. Machine Learning Workflow By Steps 3. Four Groups Of Task That Machine Learning Solves 3.1 Classification 3.2 Cluster analysis 3.3 Regression 3.4 Ranking 3.5 Generation 4. Three Model Training Styles 4.1 Supervised learning 4.2 Unsupervised learning 4.3 Reinforcement learning 5. Embarking On Machine Learning 5.1 "Business translator" and .