Exploring the Construction of an Entertainment Brain: AI and Big Data Practices in the Fish Brain Platform

This article records a talk from the Youku Technology Salon where senior Alibaba engineer Mu Ji presented "Exploring the Construction of an Entertainment Brain". The talk focuses on the Fish Brain platform, an AI‑driven decision‑support system designed for the entertainment industry.

The speaker first compares the Fish Brain platform with well‑known large‑scale AI platforms such as Google Brain and Baidu Brain, noting that while those focus on general machine perception and natural language processing, Fish Brain is tailored to the specific needs of content creation, acquisition, production, and distribution.

Because the entertainment content market has exploded in volume and diversity, traditional manual analysis is insufficient. The platform therefore aims to provide auxiliary decision‑making by integrating rich content data, user behavior, and contextual information.

Fish Brain adopts a three‑layer architecture:

Data‑model layer: crawls and processes all online entertainment content, storing it in a foundational data warehouse.

Basic‑data layer: applies AI algorithms (sentiment classification, user‑psychology tagging, face recognition, video exposure detection, beauty scoring, etc.) to generate intermediate data.

Application‑model layer: consumes the processed data to support downstream products.

The talk then dives into three technical pillars of the middle layer:

Natural Language Processing (NLP) : Uses sequence models (RNN, LSTM) and embeddings (Word2vec, BERT) to build universal language models. Emphasis is placed on the generalization power of embedding‑based representations and on incorporating domain knowledge for script and IP analysis. A case study on the classic novel "The Legend of the Condor Heroes" demonstrates character‑level conflict detection, relationship strength analysis, and storyline conflict curves.

Computer Vision (CV) : Discusses the limitations of generic embeddings for visual tasks and presents custom architectures (e.g., TextCNN‑style networks, specialized loss functions such as SoftMax, Margin loss, Focal loss, and SVM‑margin loss). Examples include facial expression recognition for emotion curve extraction and genre‑specific emotion density analysis.

Predictive Capability : Describes a hybrid model combining DNN, Relation Net, and Multi‑Task Learning (MTL) with transfer learning to predict the performance of unreleased content. Data augmentation and uncertainty learning (variational Bayesian networks) are employed to mitigate data scarcity and improve prediction accuracy by 3‑4%.

In summary, the Fish Brain platform integrates AI and big‑data techniques—NLP, CV, embedding, loss design, uncertainty modeling—to provide comprehensive decision support for the entertainment industry.