BCA 314T Deep Learning with Python Notes cover all 4 units of this GGSIPU BCA Semester 6 elective in 72 pages: 14 Keras and NumPy programs, 31 diagrams, 49 solved two-mark answers and 33 ten-mark questions. They follow the scheme for the 2024-25 batch onwards and come as a printed book.
Written by Prateek Talwar, Founder, Unnati Educations · Reviewed by Ritik Rawat, B.Tech, Senior Full Stack Developer · Checked against the GGSIPU BCA Scheme of Examination and Syllabi, 08 August 2026 · Updated 13 October 2026 · Editorial policy
Scheme: BCA scheme 2024-25 (USICT) | Applies to: batch admitted 2024-25 onwards | Verified against: official BCA scheme PDF, 08 August 2026
BCA 314T Deep Learning with Python at a glance: GGSIPU BCA Sem 6 credits, units and exam pattern
BCA 314T is one of four Sem 6 electives; you take one with its own lab. It sits with the other GGSIPU BCA Semester 6 papers in the 2024-25 scheme.
| Code and title | BCA 314T · Deep Learning with Python |
|---|---|
| Course and semester | BCA · Semester 6 |
| Category and credits | Programme Core Elective (PCE) · 4 credits · L-T-C 4-0-4 |
| Marks | 100: 40 internal + 60 end-term |
| Applies to | Batch admitted in 2024-25 onwards |
| End-term pattern | Q1: 10 compulsory parts of 2 marks; then one 10-mark question per unit, from 2 choices (20 + 40 = 60 marks) |
| Exam duration | 3 hours |
| Units | 4 (11, 12, 11 and 11 lecture hours) |
| Paired lab | BCA 314P Deep Learning with Python Lab · PCE · 2 credits · 100 marks (40 + 60) |
Outcomes CO5 and CO6 ask for an RNN and object detection, so the GGSIPU BCA Sem 6 Deep Learning with Python notes give Unit 4 an LSTM and a YOLO program.
BCA 314T Deep Learning with Python Notes unit-wise: 14 Python programs and 31 diagrams across 4 units
These BCA 314T Deep Learning with Python IPU notes follow the official unit order; the PDF leaves units untitled, so headings use the notes' titles. Other papers: GGSIPU BCA notes.
One page per unit is in the sample PDF above.
BCA 314T Unit 1: Neural Network Foundations
Pages 3 to 22. The perceptron rule and its XOR failure, activation and loss functions, gradient descent against SGD, backpropagation with a one-neuron numerical, dropout and batch normalisation, ending in a Keras ANN on MNIST with 109,898 parameters counted layer by layer. 14 sections, 2 programs, 9 figures, 12 two-mark answers, 8 ten-mark questions.
Quick look: a perceptron for the AND gate (from page 8)
| x1, x2 | z = x1 + x2 − 1.5 | step(z) |
|---|---|---|
| 0, 0 | −1.5 | 0 |
| 0, 1 or 1, 0 | −0.5 | 0 |
| 1, 1 | +0.5 | 1 |
Quick look: AND gate with w1 = w2 = 1, b = −1.5 (page 8)
- x1, x2
- 0, 0
- z = x1 + x2 − 1.5
- −1.5
- step(z)
- 0
- x1, x2
- 0, 1 or 1, 0
- z = x1 + x2 − 1.5
- −0.5
- step(z)
- 0
- x1, x2
- 1, 1
- z = x1 + x2 − 1.5
- +0.5
- step(z)
- 1
BCA 314T Unit 2: Convolutional Neural Networks
Pages 23 to 39. Why dense layers waste weights on images (9.6 million against 1,792 for 64 filters), then convolution, padding, pooling and softmax with cross-entropy, a 225,034-parameter Keras CNN, LeNet-5 to DenseNet in one table, and He initialisation. 11 sections, 5 programs, 7 figures, 12 two-mark answers, 9 ten-mark questions.
Quick look: convolution output size (from page 27)
| W, F, P, S | Working | O |
|---|---|---|
| 28, 3, 0, 1 | (28 − 3) / 1 + 1 | 26 |
| 32, 5, 2, 1 | (32 − 5 + 4) / 1 + 1 | 32 |
| 227, 11, 0, 4 | (227 − 11) / 4 + 1 | 55 |
Quick look: O = (W − F + 2P) / S + 1 (page 27)
- W, F, P, S
- 28, 3, 0, 1
- Working
- (28 − 3) / 1 + 1
- O
- 26
- W, F, P, S
- 32, 5, 2, 1
- Working
- (32 − 5 + 4) / 1 + 1
- O
- 32
- W, F, P, S
- 227, 11, 0, 4
- Working
- (227 − 11) / 4 + 1
- O
- 55
BCA 314T Unit 3: DBN, Autoencoders and Reinforcement Learning
Pages 40 to 55. RBMs and contrastive divergence lead into Deep Belief Networks, then PCA and three Keras autoencoders (simple, denoising, deep 784-128-64-32-64-128-784). Reinforcement learning covers MDP and a Q-learning corridor program with output. 8 sections, 4 programs, 7 figures, 12 two-mark answers, 8 ten-mark questions.
Quick look: three kinds of learning (from page 53)
| Basis | Supervised | Unsupervised | Reinforcement |
|---|---|---|---|
| Feedback | Correct answer per input | None | Reward, often delayed |
| Goal | Predict labels | Find hidden structure | Maximise total reward |
Quick look: learning types (page 53)
- Basis
- Feedback
- Supervised
- Correct answer per input
- Unsupervised
- None
- Reinforcement
- Reward, often delayed
- Basis
- Goal
- Supervised
- Predict labels
- Unsupervised
- Find hidden structure
- Reinforcement
- Maximise total reward
BCA 314T Unit 4: RNN, LSTM, Object Detection and Transfer Learning
Pages 56 to 72. RNN equations and Backpropagation Through Time, LSTM and GRU (gate maths in the excerpt below), sliding-window forecasting with a Keras LSTM, IoU, R-CNN, YOLO and SSD with a YOLO program, and MobileNetV2 transfer learning. 6 sections, 3 programs, 8 figures, 13 two-mark answers, 8 ten-mark questions.
Quick look: one RNN time step (from page 58)
| Step | Calculation | Result |
|---|---|---|
| Weighted sum | 0.8 × 1 + 0.4 × 0.5 + 0.1 | 1.1 |
| Hidden state | tanh(1.1) | 0.8005 |
| Output | 1.5 × 0.8005 | 1.2007 |
Quick look: x = 1, previous h = 0.5, Wxh = 0.8, Whh = 0.4, b = 0.1, Why = 1.5 (page 58)
- Step
- Weighted sum
- Calculation
- 0.8 × 1 + 0.4 × 0.5 + 0.1
- Result
- 1.1
- Step
- Hidden state
- Calculation
- tanh(1.1)
- Result
- 0.8005
- Step
- Output
- Calculation
- 1.5 × 0.8005
- Result
- 1.2007
BCA 314T exam corner: 49 solved two-mark and 33 ten-mark questions by unit
Each unit ends with answered two-mark questions and ten-mark questions. No year tags appear, so treat them as exam-style questions.
| Unit | 2 Marks (answered) | 10 Marks | Programs |
|---|---|---|---|
| 1 | 12 | 8 | 2 |
| 2 | 12 | 9 | 5 |
| 3 | 12 | 8 | 4 |
| 4 | 13 | 8 | 3 |
BCA 314T questions and programs in the notes, by unit
- Unit
- 1
- 2 Marks (answered)
- 12
- 10 Marks
- 8
- Programs
- 2
- Unit
- 2
- 2 Marks (answered)
- 12
- 10 Marks
- 9
- Programs
- 5
- Unit
- 3
- 2 Marks (answered)
- 12
- 10 Marks
- 8
- Programs
- 4
- Unit
- 4
- 2 Marks (answered)
- 13
- 10 Marks
- 8
- Programs
- 3
- Unit 1: What is dropout? Random neurons switch off in each training step, never at test time. Quote
Dropout(0.2). - Unit 2: Why is zero weight initialisation not used? Every neuron gets the same gradient, so none learns a different feature. Name He initialisation.
- Unit 3: What is the discount factor γ? A value from 0 to 1 that shrinks future rewards: 1, 1, 1 with γ = 0.9 give 2.71.
- Unit 4: What is Non-Maximum Suppression? Keep the top-scoring box; drop overlapping boxes above an IoU threshold.
The BCA 314T notes solve 49 two-mark questions and set 33 ten-mark questions across 4 units, plus 14 Python programs. More practice sits in the BCA 314T important questions.
Who writes and checks the BCA 314T Deep Learning notes?
- Prateek Talwar, author: Founder, Unnati Educations; ACS, ICSI (membership no. 58704, since 15 May 2019); B.Com, University of Delhi (SOL), 2012. Profile
- Ritik Rawat, subject reviewer: B.Tech; Senior Full Stack Developer, Unnati Educations. He checks the 14 Keras and NumPy programs and the parameter tables. Profile
The 2026-27 edition is mapped to the GGSIPU BCA scheme (Table 11, Sixth Semester), last updated 13 October 2026, under our editorial policy.
The same pair prepares GGSIPU important questions, which for BCA 314T will add CNN architecture and LSTM practice.
How to write BCA 314T answers: formula-first theory, Keras code and architecture diagrams
Deep learning answers lose marks in two ways: a definition with no formula, or code with no explanation. The exam tips inside the BCA 314T notes fix both, and each tip names exactly what a full answer must contain.
- Activation functions (page 11): formula, output range, one advantage, one disadvantage, then the rule: ReLU in hidden layers; sigmoid, softmax or linear at the output.
- Backpropagation (page 14): name forward pass, loss, chain rule, backward pass and weight update, in that order.
- CNN architectures (page 34): authors and year, ILSVRC result, layers and parameters, a block diagram and three or four key ideas. Draw the residual block or inception module.
- Deep Belief Networks (page 44): Hinton, Osindero and Teh, 2006; greedy layer-wise pre-training plus fine-tuning; a stacked-RBM diagram.
- Transfer learning (page 69): link it to detection, since Faster R-CNN, SSD and YOLO use ImageNet-trained backbones.
For a ten-mark program, follow the five steps every Keras program in the notes uses: load and scale data, build, compile(), fit(), evaluate(). Then add an output-shape and parameter table, and one line on why you chose each activation and loss, for example softmax with sparse_categorical_crossentropy because MNIST labels are integers 0 to 9.
Worked example: one Conv2D row
The second Conv2D(64, (3, 3)) layer of the MNIST CNN gets a 13 × 13 × 32 input. Output size is (13 − 3) / 1 + 1 = 11, so the shape is 11 × 11 × 64, and parameters are (3 × 3 × 32 + 1) × 64 = 18,496.
Two-mark answers stay at two or three lines with one formula or example, like the 49 model answers in the notes. For numericals such as the 32 × 32 input with a 5 × 5 filter, write the formula before the numbers. Practise this format on GGSIPU BCA important questions.
BCA 314T and the BCA 314P Deep Learning lab (teacher-set practical list), plus other GGSIPU BCA Sem 6 papers
Paired lab: BCA 314P Deep Learning lab notes. 2 credits, 100 marks. Your teacher notifies the practical list in week one, so there is no fixed count; the theory programs are the natural starting point:
- Unit 1 → NumPy perceptron, Keras ANN
- Unit 2 → Keras CNN, VGG16, Keras Tuner
- Unit 3 → three autoencoders, Q-learning
- Unit 4 → LSTM forecasting, YOLO, MobileNetV2
Other Sem 6 papers: BCA 302T Distributed Systems and Cloud Computing, BCA 304T Machine Learning (closest link, as Unit 1 opens with an ML overview), BCA 306T Software Project Management and BCA 320 Minor Project. Alternative electives: BCA 312T, 316T and 318T.
LSTM gates and the cell state in deep learning: excerpt from page 60 of the BCA 314T notes
Excerpt from page 60 (Unit 4, Section 3: Long Short-Term Memory)
The six LSTM equations
An LSTM cell joins the previous hidden state and the current input as [h(t−1), x(t)] and passes them through three gates and one candidate layer. Each gate uses the sigmoid σ, so its output lies between 0 and 1. The symbol ⊙ means element-wise multiplication, applied position by position.
- Forget gate: f(t) = σ(Wf · [h(t−1), x(t)] + bf)
- Input gate: i(t) = σ(Wi · [h(t−1), x(t)] + bi)
- Candidate: C̃(t) = tanh(WC · [h(t−1), x(t)] + bC)
- Cell state: C(t) = f(t) ⊙ C(t−1) + i(t) ⊙ C̃(t)
- Output gate: o(t) = σ(Wo · [h(t−1), x(t)] + bo)
- Hidden state: h(t) = o(t) ⊙ tanh(C(t))
What each gate decides
| Gate | Question it answers | Example in a language model |
|---|---|---|
| Forget gate f(t) | Which old information should leave the cell state? | A new subject appears, so forget the old subject's gender and number |
| Input gate i(t) with candidate C̃(t) | Which new information should be stored? | Store the gender and number of the new subject |
| Output gate o(t) | Which part of the memory should be output now? | Output what is needed to choose the next verb form |
The question each LSTM gate answers, with a language-model example (page 60)
- Gate
- Forget gate f(t)
- Question it answers
- Which old information should leave the cell state?
- Example in a language model
- A new subject appears, so forget the old subject's gender and number
- Gate
- Input gate i(t) with candidate C̃(t)
- Question it answers
- Which new information should be stored?
- Example in a language model
- Store the gender and number of the new subject
- Gate
- Output gate o(t)
- Question it answers
- Which part of the memory should be output now?
- Example in a language model
- Output what is needed to choose the next verb form
Why the cell state fixes vanishing gradients
In a plain RNN, the gradient is multiplied by the same weight matrix at every time step, so over a long sentence it shrinks towards zero. The LSTM cell state is different: it is updated mainly by addition, C(t) = f(t) ⊙ C(t−1) plus the new information.
When the forget gate stays close to 1, the gradient flows backwards along the cell state almost unchanged. This "gradient highway" works much like the skip connection in ResNet, and it lets an LSTM remember information across hundreds of time steps. That is why LSTMs beat simple RNNs on long sentences, speech and long time series.
Do the BCA 314T notes cover the full GGSIPU syllabus?
Yes. Every topic in the four official units has a section in the notes; Unit 3 also adds the Restricted Boltzmann Machine before Deep Belief Networks.
| Official unit | Official topics | Where in the notes | Status |
|---|---|---|---|
| Unit I | Machine Learning overview, ANN, Perceptron, training a Neural Network, Activation Functions, Loss Function, Hyperparameters, Gradient Descent, Stochastic Gradient Descent, Backpropagation and regularization, Batch normalization, Building an ANN in Python, TensorFlow, Keras | Sections 1 to 14, pages 4 to 20 | Covered |
| Unit II | Deep vs Shallow Networks, CNN: Convolution Layers, Pooling, Flattening, Fully Connected Layers, Softmax and Cross-Entropy, CNN in Python, Fully Connected CNN, LeNet, AlexNet, ZFNet, GoogLeNet, VGGNet, ResNet, DenseNet, weights initialization, batch normalization, hyperparameter optimization | Sections 1 to 11, pages 24 to 37 | Covered |
| Unit III | Deep Belief Networks, Dimensionality Reduction, Autoencoder, Denoising Autoencoders, Deep Autoencoders, Reinforcement Learning | Sections 1 to 8, pages 41 to 53 | Covered |
| Unit IV | RNN, LSTM, Sequence Prediction and Time Series Forecasting with LSTM, Object Detection Techniques, Transfer Learning | Sections 1 to 6, pages 57 to 69 | Covered |
Official BCA 314T units mapped to the notes
- Official unit
- Unit I
- Official topics
- Machine Learning overview, ANN, Perceptron, training a Neural Network, Activation Functions, Loss Function, Hyperparameters, Gradient Descent, Stochastic Gradient Descent, Backpropagation and regularization, Batch normalization, Building an ANN in Python, TensorFlow, Keras
- Where in the notes
- Sections 1 to 14, pages 4 to 20
- Status
- Covered
- Official unit
- Unit II
- Official topics
- Deep vs Shallow Networks, CNN: Convolution Layers, Pooling, Flattening, Fully Connected Layers, Softmax and Cross-Entropy, CNN in Python, Fully Connected CNN, LeNet, AlexNet, ZFNet, GoogLeNet, VGGNet, ResNet, DenseNet, weights initialization, batch normalization, hyperparameter optimization
- Where in the notes
- Sections 1 to 11, pages 24 to 37
- Status
- Covered
- Official unit
- Unit III
- Official topics
- Deep Belief Networks, Dimensionality Reduction, Autoencoder, Denoising Autoencoders, Deep Autoencoders, Reinforcement Learning
- Where in the notes
- Sections 1 to 8, pages 41 to 53
- Status
- Covered
- Official unit
- Unit IV
- Official topics
- RNN, LSTM, Sequence Prediction and Time Series Forecasting with LSTM, Object Detection Techniques, Transfer Learning
- Where in the notes
- Sections 1 to 6, pages 57 to 69
- Status
- Covered
Source: GGSIPU BCA Scheme of Examination and Syllabi 2024-25.
Every course in GGSIPU notes is mapped to its official PDF the same way.
BCA 314T Deep Learning notes FAQs
How do I get the full BCA 314T notes?
Tap "Get Full Notes" at the top or the WhatsApp box at the end. The message opens with the paper code and a reference code already typed. The full BCA 314T notes are sold only as a 72-page printed book, and our team confirms your order and the dispatch details for your address on WhatsApp.
Are the BCA 314T notes available as a PDF?
No. The full BCA 314T notes come only as a printed book. The one PDF is the watermarked sample behind "View Sample PDF", with the cover, the contents page and one page from each unit, so you can judge the diagrams, tables and Keras code layout yourself before you order the printed book on WhatsApp.
How is BCA 314T marked, and what is the passing mark?
BCA 314T is marked out of 100: 40 internal and 60 end-term. The end-term paper has ten compulsory 2-mark parts (20 marks) and four 10-mark questions, one per unit (40 marks). The BCA scheme PDF prints no pass percentage for this paper, so check end-term minimums and re-appear rules in GGSIPU Ordinance 11 on ipu.ac.in.
Is BCA 314T compulsory in BCA Semester 6?
No. BCA 314T is a Programme Core Elective. In Semester 6 you choose one elective theory paper with its lab: BCA 312T, 314T, 316T or 318T. If you take BCA 314T, you also take BCA 314P Deep Learning with Python Lab for 2 credits. Which electives run depends on your college, so confirm with your department.
Which BCA 314T unit has the most Python programs?
Unit 2, with 5 of the 14: the Keras CNN on MNIST, single-image prediction, pre-trained VGG16 and ResNet50, a Conv, batch normalisation and ReLU block, and a Keras Tuner search. Unit 3 has 4, Unit 4 has 3 and Unit 1 has 2. The NumPy perceptron and the Q-learning corridor also print their output in the notes.
Do I need BCA 304T Machine Learning before BCA 314T?
No. The official BCA 314T syllabus lists one pre-requisite: basics of Python programming. Unit 1 of the BCA 314T notes opens with a short machine learning overview, covering the types of learning, the ML workflow and why deep learning removes manual feature engineering, so you can start even while BCA 304T runs in the same semester.
Get the full BCA 314T Deep Learning with Python Notes on WhatsApp
Sold only as a printed book, not a PDF, and ordered on WhatsApp.
| Unit | Pages |
|---|---|
| 1. Neural Network Foundations | 3 to 22 (20) |
| 2. Convolutional Neural Networks | 23 to 39 (17) |
| 3. DBN, Autoencoders and RL | 40 to 55 (16) |
| 4. RNN, LSTM, Detection and Transfer Learning | 56 to 72 (17) |
BCA 314T printed book: pages per unit
- Pages
- 3 to 22 (20)
- Pages
- 23 to 39 (17)
- Pages
- 40 to 55 (16)
- Pages
- 56 to 72 (17)
Format: Deep Learning with Python notes in English, a 72-page printed book with 14 programs, 49 two-mark answers, 33 ten-mark questions and 31 figures.
Order the BCA 314T Deep Learning with Python Notes book on WhatsApp (4 units, 72 pages)
Unnati Educations is an independent education provider based in Vikaspuri, Delhi. We are not affiliated with, authorised by, or endorsed by Guru Gobind Singh Indraprastha University (GGSIPU). 'GGSIPU' and 'IPU' are used only to describe the syllabus our notes follow. For official information, visit ipu.ac.in.
Ready for the full notes?
Discuss coverage and delivery with our team.