Deep Learning: NLP for Sentiment analysis & Translation 2025

Learn to initialize and cast tensors in TensorFlow, creating 0d to 4d tensors, exploring data types, shapes, and printing results, and generating random, uniform, and identity tensors for deep learning.

Explore tensor indexing in TensorFlow, learning to access 1D, 2D, and 3D tensors using indices, slices, negative indices, and step sizes to extract values, ranges, and specific patterns.

Explore linear algebra operations in TensorFlow, including matrix multiplication with matmul, transpose, 2d/3d tensors, insum, dot and outer products, and band part techniques.

Explore common TensorFlow methods such as expand_dims, squeeze, reshape, concatenate, stack, gather, and gather_nd, and learn how to use padding and broadcasting to shape tensors.

Explore ragged tensors and learn why non-rectangular data require them, create ragged tensors with tf.ragged.constant, and use boolean masks and row splits for variable-length sequences.

Learn how sparse tensors in TensorFlow use indices, values, and a dense shape to efficiently represent tensors with many zeros and convert to dense.

Explore how TensorFlow handles strings with the TensorFlow strings module, creating 1D string tensors, performing join operations, and converting inputs to strings for downstream tasks.

Create and manipulate TensorFlow variables by initializing with a tensor, fixing type and shape, and updating values with assign or assign add methods.

Prepare data for deep learning by loading the second hand cars dataset, selecting features, creating inputs and outputs, shuffling to prevent bias, and applying normalization with TensorFlow.

Build a simple linear regression model in TensorFlow using weights and biases, via a sequential API with normalization and dense layers, and study model summary and trainable parameters.

Explore how to train and optimize a simple model using stochastic gradient descent, adjust weights with learning rate, monitor loss across epochs, and consider Adam as a popular alternative.

Learn to split data into train, validation, and test sets, prevent data leakage with proper normalization, and monitor training versus validation performance using plots and metrics.

Address underfitting by increasing model complexity, adding deeper layers and non-linear activations, and monitoring training and validation loss to improve performance in deep learning models.

Explore how the TensorFlow data API builds efficient data processing pipelines with from slices method, shuffle, batch, and prefetch to handle large datasets for training, validation, and testing.

Explore how sentiment analysis derives opinions from written text and classifies reviews as positive or negative. Learn preprocessing steps: standardization, tokenization, encoding for IMDB data in deep learning.

Standardize text for NLP by converting to lowercase, removing HTML tags and punctuation, then apply stemming or lemmatization before mapping words to numbers via a vocabulary.

Explore character, word, subword, and n-gram tokenization for nlp, build vocabularies, map tokens to IDs, and use one-hot encoding to prepare model inputs.

Learn how tf-idf weights words using term frequency and inverse document frequency to identify meaningful tokens in sentences and the dataset.

Explore how embeddings convert a large vocabulary of one-hot vectors into dense, 300-dimensional word vectors, enabling semantic relationships where similar words cluster and the embedding matrix learns from data.

Explore advanced recurrent neural networks, including LSTM and GRU, and learn how to tackle vanishing gradients, short-term memory, and bi-directional architectures for sentiment analysis.

Train a sentiment analysis model using a pre-trained word2vec embedding loaded via gensim, mapping a 10,000-word vocabulary to 300-dimensional vectors from a 3 million word space, with trainable embedding layer.

Test a sentiment analysis model by feeding external text, observe how vectorization and embeddings influence predictions, and deploy an inference-ready model with an integrated vectorization layer.

Visualize word embeddings with Tensorboard by projecting 300-dimensional vectors to a 3D space using PCA or t-sne, aided by a projector, metadata, and distance metrics.

Explore building a neural machine translation model with TensorFlow, using an English-French dataset, shuffling data, and vectorizing text with separate English and French vocabularies.

Implement a custom BLEU score metric in TensorFlow from scratch, converting predictions to vector form, applying a boolean mask to count matches, and evaluating on translation data.

Discover Bahdanau attention, which links each output to all input positions via attention weights, overcoming the fixed-length bottleneck in encoder-decoder neural machine translation.

Build and test Bahdanau attention in a seq2seq translator by constructing an encoder with embedding and lstm, an attention layer, and a gru decoder, plus evaluation with bleu scores.

Understand the transformer, a breakthrough NLP model that uses attention only, enabling parallelization. Grasp self-attention, multi-head attention, encoder-decoder architecture, positional encoding, and training benefits.

Build, train, and test transformers from scratch, implementing positional encoding and embeddings, encoder-decoder attention with padding, causal and cross masks, and a scalable training schedule.

Build a custom multi-head self-attention layer from scratch, including masking and softmax, and integrate it into a transformer encoder–decoder. Compare parameter counts with TensorFlow and demonstrate translation tasks in NLP.

Visualize transformer attention by extracting and plotting multi-head attention scores across encoder–decoder layers to show how each input word attends to others, using the start token and masking.

Apply a transformer encoder to sentiment analysis using embeddings, positional encoding, and a sigmoid dense classifier; train and evaluate with configurable layers to reach about 85% test accuracy.

Discover how locality sensitive hashing attention enables memory-efficient transformers for long sequences, using chunking, bucketing, and look-back attention.