Huawei CBG AI Challenges Parkhomenko Denis, Bankevich Sergey, Korikov Kirill, Bakarov Amir
Huawei CBG AI Challenges Computer Vision Challenges Speech & Language Challenges NLU ASR
How to make neural net lighter? State-of-the-art neural nets are very complex in terms of - calculation - size How to incorporate them in so small chips? ImageNet 1K validation set accuracy
How to make neural net lighter? Tensor decomposition: Filter quantization, dictionary based convolutions: Target platform optimization: - deep knowledge of CPU/TPU architecture - vectorization, intrinsics, code optimization R&D in new methods of matrix, tensors decomposition Optimal parameter search Consider functional spaces C1(X,Y), C2(X,Y). For any given model params θ=(θ1,…, θN) and model f1(x,θ)∈C1 find f2∈C2 such that: 𝑓 2 (x)= 𝑎𝑟𝑔𝑚𝑖𝑛 𝑓 2 ∈ 𝐶 2 ( 𝐸 𝑥~𝑋 || 𝑓 1 (𝑥,𝜽)− 𝑓 2 (𝑥)||) If you good at low-level programming
Optical character recognition Task 1: Text detection in image Task 2: Text recognition in cropped image Task 3: End-to-end detection+recognition Task 4: Inpaiting Humans are still better there [https://towardsdatascience.com/image-inpainting-humans-vs-ai-48fc4bca7ecc]
Huawei CBG AI Challenges Computer Vision Challenges Speech & Language Challenges NLU ASR
Dialogue System
Amazon Alexa Skill Builder Interface Intent Detection Amazon Alexa Skill Builder Interface
Whole-sentence features Intent Detection Corpus Word vector Neural network Whole-sentence features Rule-based heuristics ... Word-level features: Entities information Syntax parsing features Word-level vectors softmax More networks External information: Dictionary Sentence-level vectors Intent Classification Network selection: RNN CNN Attention / Transformer … Sentence vector
NLU Challenges Deep learning model needs a lot of labeled data For our skills we could use assessors to generate and classify corpus But for third-party skills we could rely only on provided corpus (usually, tens of samples) Is it possible to build a good classifier using such a small amount of data?
Challenge Deep learning model needs a lot of labeled data For our skills we could use assessors to generate and classify corpus But for third-party skills we could rely only on provided corpus (usually, tens of samples) Is it possible to build a good classifier using such a small amount of data? Example: 10 samples of labeled data 100 samples of unlabeled data train the model on 100 samples and transfer labels
More challenges Word sense disambiguation Cross-lingual transfer Integration of knowledge graphs to supervised models Anaphora and coreference resolution Chit-chatting support Personalization of conversational agents …
Huawei CBG AI Challenges Computer Vision Challenges Speech & Language Challenges NLU ASR
ASR task Convert audio input to text output Applications Voice assistants (phone, home, car) Recording/voice input transcription Movie captions
ASR pipeline Feature extraction Acoustic model: morphemes/letters Language model, decoder: text Postprocessing
ASR components Support specific input conditions Language, accent Close/far field Deal with noise, multiple people speaking, low volume/quality Different hardware
ASR components Support specific input conditions Provide specific output properties Normalization Domains
ASR components Support specific input conditions Provide specific output properties Related and relative tasks Voice activity detection Trigger phrase Direct classification
ASR challenges Speaker diarization, cocktail party, denoise Flexible language model Handling variety of accents ASR on device Text normalization Optimization for production: C/C++, low-level