Literature Review
Open in the Padhaiko app3.1 Basic literature review and background study
A literature review and background study establishes the technical foundation of your Major Project before you lock down your architecture or run experiments. Examiners look at this section to verify two things: (1) you are not reinventing solved problems blindly, and (2) your proposed design or hypothesis is grounded in verified prior work rather than guesswork.
Literature Review and Background Study
A literature review is a critical, structured synthesis of peer-reviewed papers, technical standards, RFCs, official documentation, and prior project reports relevant to a research or engineering problem. Unlike an annotated list that merely summarizes one paper after another, a proper review groups approaches by theme, compares trade-offs, and isolates the specific gap the Major Project addresses. The background study provides the broader domain context—problem prevalence, standard terminology, regulatory or operational constraints—needed for a reader to understand the rest of the report.
3.1.1 Systematic Workflow for Literature Search
For a seventh-semester B.E. Major Project, aiming for 12 to 20 high-quality, directly relevant sources (mostly published within the last 5–7 years, plus foundational seminal papers) is far better than padding the bibliography with 50 unread links. Follow a repeatable five-step method:
- Break the topic into search concepts: Separate domain (e.g., Nepali OCR), technique (e.g., transformer or CRNN), and constraint (e.g., degraded scans, mobile latency).
- Query reliable databases: Search IEEE Xplore, ACM Digital Library, SpringerLink, ScienceDirect, Google Scholar, arXiv (check whether a paper is peer-reviewed), and Institute of Engineering/Pulchowk or campus thesis archives.
- Apply inclusion/exclusion screening: Skim titles and abstracts first. Exclude papers outside your target constraint (for instance, models needing 80 GB GPUs when you target edge deployment) unless used only as an upper-bound reference.
- Extract structured notes into a matrix: Record Author/Year, Problem & Dataset, Method/Architecture, Evaluation Metric & Score, and Stated Limitations for every paper.
- Synthesize by theme, not chronologically alone: Group works by approach family (e.g., Rule-Based vs. Classical ML vs. Deep Learning) and end with a crisp Gap Summary linking directly to your Project I objectives.
| Source Type | Reliability Level | Best Used For | Caution |
|---|---|---|---|
| Peer-reviewed journals & conferences (IEEE, ACM, NeurIPS, CVPR, USENIX) | High | Benchmark methods, experimental baselines, proven limitations | Verify dataset realism and evaluation split validity |
| Standards, RFCs, & Official Docs (IETF, W3C, IEEE, ISO, NIST) | High | Protocol rules, security standards, API/hardware specs | Cite exact version or RFC number |
| Preprints (arXiv, bioRxiv) | Medium (unreviewed) | Emerging recent architectures or benchmarks | Cross-check if later published; inspect reproducibility |
| Past Institute of Engineering Major/Minor Project Reports | Context-specific | Local benchmarks, campus/Nepali datasets, avoiding repeated pitfalls | Never treat student reports as primary scientific proof |
| Blogs, Medium posts, random GitHub READMEs | Low for academic citations | Quick implementation tips or tooling notes | Do not cite as authoritative literature in formal reports |
3.2 Study of relevant theories
While Section 3.1 examines who solved similar problems and how, Study of Relevant Theories (often forming the core of the Background / Theoretical Framework section) explains the underlying mathematical, algorithmic, architectural, or protocol foundations your project builds upon.
Study of Relevant Theories
The study of relevant theories covers the core scientific and engineering principles—algorithms, mathematical models, network protocols, hardware interfaces, or software engineering patterns—required to justify why your chosen methodology works and under what assumptions it holds.
3.2.1 Selecting Only Relevant Theory (Avoiding Textbook Padding)
A frequent mistake in Major Project reports is pasting 15 pages of introductory textbook material (for example, explaining "What is a neuron?" or "What is HTTP?") that examiners skip. Instead, include only the theory that directly informs your design decisions or experimental analysis:
| Project Domain | What to Skip (Too Elementary) | What to Study & Explain Deeply |
|---|---|---|
| AI / Computer Vision / NLP | Generic history of AI, basic definition of supervised learning | Loss functions (e.g., CTC loss, Focal loss), attention/convolution mechanics relevant to your chosen model, evaluation metrics (Precision, Recall, F1, mAP, BLEU) |
| Networks & Cybersecurity | OSI 7-layer textbook diagram from scratch | Specific protocol state machines (e.g., TLS 1.3 handshake, MQTT QoS levels), cryptographic primitives (AES-GCM, Ed25519), threat model (STRIDE) |
| Embedded Systems & IoT | Basic Ohm's law or pinout of a breadboard | Sensor sampling rate & Nyquist limits, ADC resolution, communication timing (I2C/SPI/UART/LoRa link budget), power-sleep state calculations |
| Web, Mobile & Big Data | Basic definition of HTML or SQL | CAP theorem / eventual consistency for offline sync, indexing strategies (B-tree vs. LSM-tree), concurrency/queueing models, RBAC security |
Synthesizing a Literature Matrix and Theoretical Study for an Offline Speech Recognition Project
A group on the Research Track is building an Offline Nepali Voice Command Recognizer for Embedded Devices. Show how they should structure both their literature review matrix (Section 3.1) and their focused study of relevant theories (Section 3.2).
Build a comparative literature matrix (not a summary paragraph list): The group collects 14 recent papers on low-resource speech recognition and keyword spotting, then summarizes key representatives in a structured table comparing acoustic front-end, model family, dataset size, word/command accuracy, and parameter/RAM footprint.
- Classic HMM-GMM systems: small footprint, weak under background street noise.
- Large multilingual transformers (e.g., Whisper-medium): strong accuracy on clean speech, exceeds embedded RAM and latency budgets.
- Depthwise-separable CNN / Conformer-tiny keyword models: run under 500 KB quantized on edge microcontrollers/SBCs, but mostly benchmarked on English SpeechCommands rather than Nepali utterances.
Derive the exact research gap: Existing high-accuracy Nepali ASR models require cloud or multi-GB GPUs, whereas tiny embedded keyword spotters lack validated Nepali phonetic/command datasets and noise-robust training under realistic room acoustics.
Select targeted theory directly tied to the pipeline:
- Acoustic Feature Extraction: Short-Time Fourier Transform (STFT), Mel-filterbank spacing, and MFCC / Log-Mel spectrograms (explaining window length, hop size, and why Mel scale matches human pitch perception).
- Depthwise Separable Convolutions & Quantization: How factorized convolutions cut multiply-accumulate operations and how INT8 post-training quantization trades tiny accuracy loss for 4x memory reduction.
- Evaluation Theory: Confusion matrix, False Rejection Rate (FRR) vs. False Acceptance Rate (FAR) trade-off, and real-time factor (RTF) on target hardware.
The group organizes prior work into a comparative literature table across model families, isolates the embedded-Nepali-noise gap, and restricts its theoretical study strictly to Log-Mel/MFCC extraction, depthwise-separable CNN quantization, and FAR/FRR/RTF evaluation metrics.
3.3 Study of existing systems
Whether you are on the Product Track, Internship-Based Track, or Research Track (when comparing against deployed tools), Study of Existing Systems analyzes real-world software products, hardware devices, open-source platforms, or current manual organizational workflows that users currently rely on.
Study of Existing Systems
A study of existing systems is a practical engineering audit of existing applications, commercial solutions, open-source tools, or manual processes that currently address (or fail to address) your target problem. Its purpose is to benchmark usability, features, cost, deployment requirements, performance, and local fit so your system design improves measurably on what already exists.
3.3.1 How to Evaluate Existing Systems Objectively
Do not simply list screenshots of three apps and write "App X is bad." Evaluate at least 3 to 5 existing systems (including the current manual process if users use paper/spreadsheets) using consistent criteria:
- Core features and workflow: What steps must a user take from input to result?
- Architecture & platform: Web, native mobile, desktop, cloud-only vs. offline-first, proprietary hardware vs. commodity sensors.
- Cost & licensing: Subscription fees, API per-call costs, or open-source license limits (GPL, MIT, Apache).
- Local context & accessibility: Nepali/Unicode support, low-bandwidth operation, power resilience, privacy/data residency.
- Observed bottlenecks: Concrete latency, accuracy, integration, or usability failures observed by actually testing the tool or interviewing users.
| Aspect | Literature Review (Academic) | Study of Existing Systems (Practical) |
|---|---|---|
| Primary Artifacts | Research papers, theses, technical benchmarks, standards | Commercial apps, open-source repos, SaaS tools, manual office/field workflows |
| Evaluation Lens | Algorithms, mathematical assumptions, dataset accuracy, computational complexity | Feature completeness, UX friction, cost, offline reliability, deployment fit |
| Output Deliverable | Baseline algorithm choice & justified research gap | Feature comparison matrix & required system improvements (SRS inputs) |
| Common Viva Question | "Which baseline paper are you beating or reproducing, and why?" | "Why won't users just use existing App X or Excel instead of your project?" |
Benchmarking Existing Systems for a Campus Equipment & Lab Inventory Portal
A Product Track team is designing a QR-Based Departmental Lab Equipment Borrowing and Maintenance Portal for an Institute of Engineering campus. Outline how they should conduct and summarize their study of existing systems.
Select representative current alternatives (including manual workflow):
- Current Manual Logbook (Department Store): Paper issue/return register kept by lab technicians.
- General Spreadsheet Templates (Excel / Google Sheets): Shared tabular tracker without barcode scanning or automated overdue alerts.
- Snipe-IT (Open-Source Asset Manager): Mature web-based IT asset platform designed for corporate offices.
Test and compare against real campus requirements:
- Paper Logbook: Zero setup cost, works during power cuts, but searching serial numbers takes minutes, damaged-component history is lost across semesters, and monthly audit reports must be tallied by hand.
- Google Sheets: Easy to start, but concurrent edits by multiple lab assistants risk accidental row overwrites, no role-based student borrowing requests, and no printable QR asset labels.
- Snipe-IT: Feature-rich for corporate IT laptops, yet heavy to self-host/customize for Institute of Engineering academic workflows (student roll-number verification, semester-end no-dues clearance, lab-bench kit bundles where 15 sensors belong to one project kit).
Summarize into a feature-gap table feeding directly into System Design:
The group highlights four verified differentiators their proposed portal provides: (1) kit-level + component-level QR tracking, (2) student roll-number authentication and overdue email/SMS notices, (3) semester-end clearance report generation, and (4) lightweight Docker deployment on campus local servers.
The team audits the current paper register, spreadsheets, and Snipe-IT across cost, workflow friction, offline/local hosting, and Institute of Engineering-specific academic rules (kit bundles and semester clearance), producing a clear feature-gap matrix that justifies building a tailored campus portal.
Viva-Style Questions: Literature Review and Background Study
What is the difference between an annotated bibliography (a paper-by-paper summary list) and a true literature review? Why do examiners criticize paper-by-paper summaries?
Annotated bibliography vs. literature review: An annotated bibliography simply lists Source 1, Source 2, Source 3 in separate paragraphs without connecting them. A literature review organizes sources by theme or methodology, compares strengths, trade-offs, and evaluation conditions across works, and builds a logical argument leading to the project's gap.
Why examiners criticize simple lists: A list shows that you collected abstracts, not that you understood how methods relate, why one approach outperforms another under your constraints, or how your design builds on prior art.
An annotated bibliography summarizes sources in isolation, whereas a literature review synthesizes them thematically, compares trade-offs in a structured matrix, and derives the exact research or engineering gap your project addresses.
Suppose your supervisor tells you: "Your Literature Review chapter has too much textbook theory and not enough research papers." How do you reorganize Chapters/Sections 3.1, 3.2, and 3.3 to fix this?
Separate prior-work synthesis from theoretical background:
- Move generic definitions out, keeping only the specific equations, protocols, or algorithmic mechanics directly used by your system inside Relevant Theories (3.2).
- Populate Literature Review (3.1) with recent peer-reviewed studies (papers/standards) that tackled your exact or adjacent problem, including their datasets, metrics, and limitations.
- Place real software tools, commercial products, and manual workflows into Study of Existing Systems (3.3) with a feature-comparison matrix.
Trim elementary textbook filler, restrict Theoretical Study (3.2) strictly to principles that govern your design and metrics, expand Literature Review (3.1) with recent peer-reviewed papers compared by method and limitations, and move product/workflow benchmarks into Study of Existing Systems (3.3).
During a viva for a Product Track project, an examiner asks: "An open-source GitHub project already does 80 percent of what you described. Why should we approve your project?" How should you answer?
Acknowledge the existing system honestly: Never pretend an existing open-source system does not exist—examiners check GitHub and Google Scholar. State clearly what the existing tool handles well.
Present verified gaps and original engineering contributions: Explain the specific technical, domain, or deployment gaps verified in your Study of Existing Systems—for example: missing support for Nepali language/local compliance formats, inability to run offline in low-connectivity clinics, lack of hardware integration with target sensors, or architectural changes you are designing and evaluating empirically.
Acknowledge the open-source tool openly as a baseline in your Study of Existing Systems, then pinpoint the concrete missing capabilities—such as offline synchronization, local language/format support, custom hardware integration, or improved latency/accuracy—that form your original engineering scope.
Why should a Major Project group avoid relying heavily on blog posts, YouTube tutorials, or unverified websites in its formal Literature Review? When is citing non-academic sources acceptable?
Limitations of informal web sources: Blogs and tutorials are not peer-reviewed, rarely report controlled baselines or failure cases, often omit dataset details, and frequently disappear or break links.
Acceptable non-academic citations: Official technical documentation (e.g., PostgreSQL docs, ESP-IDF datasheets), IETF RFCs, IEEE/ISO standards, and official government/NGO statistical reports are valid authoritative sources for specifications and domain statistics.
Informal blogs lack peer review, experimental rigor, and reproducibility. Formal literature reviews should rely on peer-reviewed papers (IEEE, ACM, etc.), official standards/RFCs, hardware datasheets, and authoritative technical documentation.
How should a Research Track team use its Literature Review to choose a baseline model and evaluation metrics before coding begins?
Selecting a reproducible baseline: Identify the simplest well-established method from your literature matrix that runs within your hardware/data constraints and has publicly known behavior. Implement that first as your baseline so every improvement you propose later is measured fairly.
Locking standard evaluation metrics: Adopt the exact metrics standard in your reviewed papers (e.g., macro-F1 for imbalanced classes rather than raw accuracy, or p95 latency for network proxies) so your Project I PoC and Project II results can be compared directly against prior work.
Use the literature matrix to pick a well-understood, hardware-feasible prior method as your reference baseline and adopt the field's standard evaluation metrics (e.g., macro-F1, mAP, or tail latency) so subsequent improvements are rigorously comparable.