Pre-Trained AI Models on Google Cloud

Pre-trained AI models are machine learning models that have been trained on large, general datasets by Google or third-party providers. They encapsulate learned patterns, features, and representations that can be reused for a variety of downstream tasks. For example, a pre-trained image classification model like ResNet-50 has already learned to recognize edges, textures, shapes, and common objects from millions of images. A pre-trained language model like BERT has learned grammar, context, and semantics from billions of words.

The primary reason pre-trained models exist is the prohibitive cost and complexity of training large models from scratch. Training a state-of-the-art model can require thousands of GPU-hours, terabytes of data, and deep expertise in machine learning. Pre-trained models democratize AI by allowing organizations to start with a powerful baseline and only fine-tune on their specific data—often with just a few hundred or thousand examples. This approach is called transfer learning.

On Google Cloud, pre-trained models are available through several services: - Vertex AI Model Garden: A curated repository of foundation models, including Google's own models (e.g., PaLM, Codey, Imagen) and open-source models (e.g., BERT, ResNet, YOLO). - Pre-built APIs: Fully managed services like Cloud Vision API, Natural Language API, Translation API, and Speech-to-Text API that provide ready-to-use inference endpoints without any model training. - AI Platform (Unified): Allows you to deploy pre-trained models from TensorFlow Hub, PyTorch Hub, or custom containers.

A pre-trained model is essentially a set of learned weights and biases that define a neural network. The training process involves two phases: 1. Pre-training: The model is trained on a large, generic dataset using unsupervised or self-supervised learning. For example, BERT is pre-trained on the entire Wikipedia corpus and BookCorpus using a masked language model objective: it learns to predict missing words in a sentence. This forces the model to understand context and syntax. 2. Fine-tuning: The pre-trained weights are used as a starting point, and the model is further trained on a smaller, task-specific dataset with a supervised objective. For instance, you can take BERT and fine-tune it on a set of customer support emails labeled by category (billing, technical, etc.). Only the final classification layer may need to be trained from scratch; the rest of the network is adjusted slightly.

During inference, the model takes input (e.g., an image, text, or audio) and passes it through its layers. The pre-trained layers extract general features (e.g., edges, shapes, syntax), and the fine-tuned layers map those features to the specific task output (e.g., category probabilities, bounding boxes, translated text).

Vertex AI Model Garden:

- Models are organized by task: text generation, image generation, code generation, vision, etc. - Foundation models include: - PaLM 2: Google's large language model for text generation, summarization, and chat. Available in sizes: Bison, Gecko. - Codey: For code generation and completion. Based on PaLM 2. - Imagen: For text-to-image generation. - Chirp: For speech recognition. - Open-source models include: BERT, ResNet, EfficientNet, YOLO, and many more from TensorFlow Hub. - Default inference endpoints are auto-scaled; you pay per prediction request. - Fine-tuning can be done via Vertex AI Pipelines or custom training jobs.

Pre-built APIs: - Cloud Vision API: Returns labels, OCR, safe search, and face detection. Default batch size: 16 images per request. Max image size: 20 MB. Supports JPEG, PNG, GIF, BMP, WEBP. - Natural Language API: Entity extraction, sentiment analysis, syntax analysis. Default max text length: 1 MB per document. - Translation API: Supports 100+ languages. Default translation model is NMT (Neural Machine Translation). - Speech-to-Text API: Supports 125+ languages. Default sample rate: 16 kHz. Can process up to 1 minute of audio per request (synchronous) or up to 480 minutes (asynchronous).

AI Platform (Unified):

Using Vertex AI Model Garden via gcloud:

gcloud ai models list --region=us-central1

This lists all models in your project. To deploy a model:

gcloud ai endpoints create --region=us-central1 --display-name=my-endpoint
gcloud ai endpoints deploy-model --region=us-central1 --endpoint=my-endpoint --model=model-id --machine-type=n1-standard-4

Using Cloud Vision API via curl:

curl -X POST -H "Authorization: Bearer $(gcloud auth application-default print-access-token)" \
  -H "Content-Type: application/json" \
  https://vision.googleapis.com/v1/images:annotate -d '{
    "requests": [{
      "image": {"source": {"imageUri": "gs://bucket/image.jpg"}},
      "features": [{"type": "LABEL_DETECTION", "maxResults": 10}]
    }]
  }'

Response includes labels with confidence scores.

Verifying model performance: Use Vertex AI Model Evaluation to compute metrics like precision, recall, and F1 score on a test set.

Pre-trained models integrate with: - Vertex AI Pipelines: Automate the fine-tuning and deployment workflow. - Cloud Storage: Store training data and model artifacts. - Cloud Functions: Trigger inference on events (e.g., new image uploaded to Cloud Storage). - BigQuery ML: Use pre-trained models directly in SQL queries for predictions. - AI Explanations: Get feature attributions for model predictions.

The GCDL exam tests understanding of when to use pre-trained vs. custom models. Key points:

Candidates often confuse pre-trained models with custom models. Common wrong answers: - "Pre-trained models always require more data than custom models" — False; they require less. - "You cannot fine-tune a pre-trained model on Google Cloud" — False; you can via Vertex AI. - "All pre-trained models are available as APIs only" — False; many are available as deployable models. - "Pre-trained models are always more accurate than custom models" — False; custom models can outperform if sufficient data and expertise exist.

Another trap: assuming that a pre-trained model cannot be used for a task different from its pre-training objective. For example, BERT (pre-trained on masked language modeling) can be fine-tuned for classification, question answering, or summarization.

A large online retailer wants to automatically categorize new product listings into 500 categories (e.g., Electronics > Laptops, Clothing > Shoes). They have 50,000 labeled products but the categories change frequently. They choose to use a pre-trained BERT model from Vertex AI Model Garden and fine-tune it on their product titles and descriptions. The fine-tuning takes 3 hours on an n1-standard-8 machine. They deploy the model to an endpoint with autoscaling (1-10 nodes). In production, the model achieves 94% accuracy. When a new category is added, they collect 100 examples and fine-tune again, which takes only 30 minutes due to transfer learning. The key benefit: they avoided building a model from scratch, saving months of development time.

A hospital network needs to extract medical codes (ICD-10) from clinical notes. They have strict data privacy requirements—data must stay within their VPC. They cannot use a public API. Instead, they deploy a pre-trained Natural Language API model in a private endpoint using Vertex AI Private Endpoints. They fine-tune the model using de-identified notes. The model is deployed in us-central1 with a single node. They process 10,000 notes per day with an average latency of 200ms. The fine-tuning improved accuracy from 70% (zero-shot) to 92%. Misconfiguration risk: if they had not used a private endpoint, data would have left their VPC, violating HIPAA.

A telecom company wants to build a chatbot to answer billing questions. They use the PaLM 2 foundation model via Vertex AI with few-shot learning: they provide 20 example conversations in the prompt. No fine-tuning is needed. The model generates responses with high relevance. They integrate with Dialogflow CX for conversation management. The system handles 100,000 conversations per month. Cost is $0.002 per prediction. Common pitfall: prompt engineering is critical—poorly structured prompts lead to irrelevant answers. They iterate on prompts using Vertex AI Prompt Optimizer.