1. Supervised Learning of Universal Sentence Representation from Natural Language Inference
@Facebook AI Research
Sentence Embedding
Sentence U and Sentence V mean the same?
BiLSTM Max pooling
Word embedding as input to produce general purpose sentence embedding
SkipThought/FastSent
Evolution of embedding per epoch so that capture more information.
Linaire classifier instead of big MLP
2.Learning weakly supervised multimodal phoneme embeddings
@ENS
Obejctif: Model children’s language learning process
McGurk Effect
Blue Lips dataset(blue to find mouth easier) Audio-visual speech corpus 16 French speakers
FFT => Obtain energy per freq
classifier to track mouth
Same network two words => same/not
Mono audio Mono videa Mono concatenation
Multi
Eval:
-
ABX Task Within Speaker Test Across Speaker Test
-
Parralisme s->t (con..) t->d (voice)
Q: Multi-task?
3. Generative Models
GAN in domains
KL Divergence: Train the generator by minimizing the distance of distribution of real data and the generated data
CNN+GAN secrets: Batch Norm: to stablize the training and help backprop Stride conv => DCGAN
VAE
sample in the middle prevent backprop
less perf than GAN
- shape of P(z) => Normal distribution (true latent variable)
- suboptimal loss
VAE-GAN competitive, not better
each dimension of z become meaningful eg. black hair, bold, etc.
infoGAN
GAN can be used as modules
IAN
semi-supervised
CycleGAN
Map unpaired images
4. CNTK
@Microsoft Machine Intelligence and Perception
Distrubuted Kernel instead of extension Project Malmo Reinforcement L Imitation L
Simulation realistic