Deep Forest Towards an Alternative to Deep Neural

  • Slides: 16
Download presentation
Deep Forest: Towards an Alternative to Deep Neural Networks Zhi-Hua Zhou and Ji Feng

Deep Forest: Towards an Alternative to Deep Neural Networks Zhi-Hua Zhou and Ji Feng LAMDA GROUP Nanjing University, China Presented by Qi-Zhi Cai

About this work http: //lamda. nju. edu. cn • Proposed an new decision tree

About this work http: //lamda. nju. edu. cn • Proposed an new decision tree ensemble method (gc. Forest) • Compare it with Deep Neural Network and get highly competitive performance • Proposed another perspective of Deep Learning http: //lamda. nju. edu. cn

Outline http: //lamda. nju. edu. cn • Introduction • gc. Forest Model – Cascade

Outline http: //lamda. nju. edu. cn • Introduction • gc. Forest Model – Cascade Forest Structure – Multi-Grained Scanning – Comparison with Deep Neural Network • Experiments • Conclution http: //lamda. nju. edu. cn

Outline http: //lamda. nju. edu. cn • Introduction • gc. Forest Model – Cascade

Outline http: //lamda. nju. edu. cn • Introduction • gc. Forest Model – Cascade Forest Structure – Multi-Grained Scanning – Comparison with Deep Neural Network • Experiments • Conclution http: //lamda. nju. edu. cn

Why Forest? http: //lamda. nju. edu. cn • Large Model Capacity • Fewer hyper-parameters

Why Forest? http: //lamda. nju. edu. cn • Large Model Capacity • Fewer hyper-parameters • Friendly for Parallel Computing http: //lamda. nju. edu. cn

Representation Learning • Distributed Representation http: //lamda. nju. edu. cn

Representation Learning • Distributed Representation http: //lamda. nju. edu. cn

Representation Learning http: //lamda. nju. edu. cn

Representation Learning http: //lamda. nju. edu. cn

Outline http: //lamda. nju. edu. cn • Introduction • gc. Forest Model – Cascade

Outline http: //lamda. nju. edu. cn • Introduction • gc. Forest Model – Cascade Forest Structure – Multi-Grained Scanning – Comparison with Deep Neural Network • Experiments • Conclution http: //lamda. nju. edu. cn

Cascade Forest Structure http: //lamda. nju. edu. cn • Diversity vs Accucracy • Use

Cascade Forest Structure http: //lamda. nju. edu. cn • Diversity vs Accucracy • Use prior probability as a supervisor • Use raw feature every layer to avoid divergence http: //lamda. nju. edu. cn

Multi-Grained Scanning http: //lamda. nju. edu. cn

Multi-Grained Scanning http: //lamda. nju. edu. cn

gc. Forest Model http: //lamda. nju. edu. cn

gc. Forest Model http: //lamda. nju. edu. cn

Comparison with DNN http: //lamda. nju. edu. cn

Comparison with DNN http: //lamda. nju. edu. cn

Experiments http: //lamda. nju. edu. cn

Experiments http: //lamda. nju. edu. cn

Influence of Multi-Grained Scannings http: //lamda. nju. edu. cn

Influence of Multi-Grained Scannings http: //lamda. nju. edu. cn

More in Experement http: //lamda. nju. edu. cn

More in Experement http: //lamda. nju. edu. cn

Conclusion http: //lamda. nju. edu. cn • We proposed the gc. Forset model –

Conclusion http: //lamda. nju. edu. cn • We proposed the gc. Forset model – Fewer hyperparameter – Competitive performance compare to DNN – Work well on different kinds of jobs http: //lamda. nju. edu. cn