General Information
Computer vision, natural language processing and personalised recommendations are just a few of the uses of artificial neural networks that are increasingly relevant to real-world problems that pose challenges for traditional data analysis techniques. This subject introduces students to the foundational ideas associated with the many variations of these models that have been developed for domains involving image data, temporal data, and natural language. This includes feed-forward, fully connected neural networks, convolutional neural networks, recurrent neural networks, and the transformer architecture. Class discussions will introduce the technical underpinnings of the models and applied sessions and assessments provide students the opportunity to experiment and apply them to a wide range of practical, real-world problems using Python.
-
Details
Academic unit: Bond Business School Subject code: DTSC13-301 Subject title: Deep Learning Through Neural Networks Subject level: Undergraduate Semester/Year: January 2025 Credit points: 10.000 -
Delivery & attendance
Timetable: https://bond.edu.au/timetable Delivery mode: Standard Workload items: - Forum: x12 (Total hours: 24) - Forum
- Personal Study Hours: x12 (Total hours: 72) - Recommended study time & reviewing materials
- Computer Lab: x12 (Total hours: 24) - Computer Lab
Attendance and learning activities: Attendance at all class sessions is expected. Students are expected to notify the instructor of any absences with as much advance notice as possible. +++++ BBS uses a self and peer-evaluation system to support students engaged in group-based assessments. Students are expected to provide this feedback in a timely fashion as part of their assessment. The information gathered is used by the educator as partial evidence of equitable contributions by all group members and helps to determine individual marks for group assessments. -
Resources
Prescribed resources: Books
- Francois Chollet (2021). Deep Learning with Python, Second Edition. n/a, Simon and Schuster 502
iLearn@Bond & Email: iLearn@Bond is the Learning Management System at Bond University and is used to provide access to subject materials, class recordings and detailed subject information regarding the subject curriculum, assessment, and timing. Both iLearn and the Student Email facility are used to provide important subject notifications.
Additionally, official correspondence from the University will be forwarded to students’ Bond email account and must be monitored by the student.
To access these services, log on to the Student Portal from the Bond University website as www.bond.edu.au
| Academic unit: | Bond Business School |
|---|---|
| Subject code: | DTSC13-301 |
| Subject title: | Deep Learning Through Neural Networks |
| Subject level: | Undergraduate |
| Semester/Year: | January 2025 |
| Credit points: | 10.000 |
| Timetable: | https://bond.edu.au/timetable |
|---|---|
| Delivery mode: | Standard |
| Workload items: |
|
| Attendance and learning activities: | Attendance at all class sessions is expected. Students are expected to notify the instructor of any absences with as much advance notice as possible. +++++ BBS uses a self and peer-evaluation system to support students engaged in group-based assessments. Students are expected to provide this feedback in a timely fashion as part of their assessment. The information gathered is used by the educator as partial evidence of equitable contributions by all group members and helps to determine individual marks for group assessments. |
| Prescribed resources: | Books
|
|---|---|
| iLearn@Bond & Email: | iLearn@Bond is the Learning Management System at Bond University and is used to provide access to subject materials, class recordings and detailed subject information regarding the subject curriculum, assessment, and timing. Both iLearn and the Student Email facility are used to provide important subject notifications. Additionally, official correspondence from the University will be forwarded to students’ Bond email account and must be monitored by the student. To access these services, log on to the Student Portal from the Bond University website as www.bond.edu.au |
Enrolment requirements
| Requisites: |
Pre-requisites:Co-requisites:There are no co-requisites |
|---|---|
| Assumed knowledge: |
Assumed knowledge is the minimum level of knowledge of a subject area that students are assumed to have acquired through previous study. It is the responsibility of students to ensure they meet the assumed knowledge expectations of the subject. Students who do not possess this prior knowledge are strongly recommended against enrolling and do so at their own risk. No concessions will be made for students’ lack of prior knowledge. Assumed Prior Learning (or equivalent):Possess demonstrable knowledge in elementary probability theory, statistics, elementary calculus and linear algebra to the level of a unit such as Quantitative Methods. |
| Restrictions: |
Nil |
Assurance of learning
Assurance of Learning means that universities take responsibility for creating, monitoring and updating curriculum, teaching and assessment so that students graduate with the knowledge, skills and attributes they need for employability and/or further study.
At Bond University, we carefully develop subject and program outcomes to ensure that student learning in each subject contributes to the whole student experience. Students are encouraged to carefully read and consider subject and program outcomes as combined elements.
Program Learning Outcomes (PLOs)
Program Learning Outcomes provide a broad and measurable set of standards that incorporate a range of knowledge and skills that will be achieved on completion of the program. If you are undertaking this subject as part of a degree program, you should refer to the relevant degree program outcomes and graduate attributes as they relate to this subject.
Subject Learning Outcomes (SLOs)
On successful completion of this subject the learner will be able to:
- Apply statistical techniques and mathematical reasoning to formulate machine learning tools for data analysis and appropriately explain key structural components and fitting algorithms.
- Design and train neural networks, including convolutional and recurrent structures and other modern extensions, and accurately identify key structural components and their significance.
- Apply neural networks to business data systems and other circumstances, such as image recognition and natural language processing.
- Articulate machine learning ideas, decisions, recommendations and other information in a clear, concise writing style tailored to a given audience.
- Demonstrate an appropriate awareness of global issues impacting decision-making paradigms and model-building exercises.
- Apply appropriate professional standards and best practices to make ethical, responsible decisions, decision-making paradigms and model-building exercises.
Generative Artificial Intelligence in Assessment
The University acknowledges that Generative Artificial Intelligence (Gen-AI) tools are an important facet of contemporary life. Their use in assessment is considered in line with students’ development of the skills and knowledge which demonstrate learning outcomes and underpin study and career success. Instructions on the use of Gen-AI are given for each assessment task; it is your responsibility to adhere to these instructions.
-
Assessment details
Type Task % Timing* Outcomes assessed Written Report MLP Model. Given a business problem and associated data, students must develop an MLP model in a technical report and communicate its performance in the context of a real-world problem for a managerial audience. 30.00% Week 5 1,2,4 Written Report CNN Model. Given an image classification problem and associated data, students must develop a CNN model in a technical report and communicate its performance in the context of a real-world problem for a managerial audience 40.00% Week 9 1,2,3,4,5,6 Written Report RNN Model. Given a problem involving sequential data, students must develop an RNN model in a technical report and communicate its performance in the context of a real-world problem for a managerial audience. 30.00% Week 13 1,2,3,4,6 - * Assessment timing is indicative of the week that the assessment is due or begins (where conducted over multiple weeks), and is based on the standard University academic calendar
- C = Students must reach a level of competency to successfully complete this assessment.
-
Assessment criteria
Assessment criteria
High Distinction 85-100 Outstanding or exemplary performance in the following areas: interpretative ability; intellectual initiative in response to questions; mastery of the skills required by the subject, general levels of knowledge and analytic ability or clear thinking. Distinction 75-84 Usually awarded to students whose performance goes well beyond the minimum requirements set for tasks required in assessment, and who perform well in most of the above areas. Credit 65-74 Usually awarded to students whose performance is considered to go beyond the minimum requirements for work set for assessment. Assessable work is typically characterised by a strong performance in some of the capacities listed above. Pass 50-64 Usually awarded to students whose performance meets the requirements set for work provided for assessment. Fail 0-49 Usually awarded to students whose performance is not considered to meet the minimum requirements set for particular tasks. The fail grade may be a result of insufficient preparation, of inattention to assignment guidelines or lack of academic ability. A frequent cause of failure is lack of attention to subject or assignment guidelines. Quality assurance
For the purposes of quality assurance, Bond University conducts an evaluation process to measure and document student assessment as evidence of the extent to which program and subject learning outcomes are achieved. Some examples of student work will be retained for potential research and quality auditing purposes only. Any student work used will be treated confidentially and no student grades will be affected.
| Type | Task | % | Timing* | Outcomes assessed |
|---|---|---|---|---|
| Written Report | MLP Model. Given a business problem and associated data, students must develop an MLP model in a technical report and communicate its performance in the context of a real-world problem for a managerial audience. | 30.00% | Week 5 | 1,2,4 |
| Written Report | CNN Model. Given an image classification problem and associated data, students must develop a CNN model in a technical report and communicate its performance in the context of a real-world problem for a managerial audience | 40.00% | Week 9 | 1,2,3,4,5,6 |
| Written Report | RNN Model. Given a problem involving sequential data, students must develop an RNN model in a technical report and communicate its performance in the context of a real-world problem for a managerial audience. | 30.00% | Week 13 | 1,2,3,4,6 |
- * Assessment timing is indicative of the week that the assessment is due or begins (where conducted over multiple weeks), and is based on the standard University academic calendar
- C = Students must reach a level of competency to successfully complete this assessment.
Assessment criteria
| High Distinction | 85-100 | Outstanding or exemplary performance in the following areas: interpretative ability; intellectual initiative in response to questions; mastery of the skills required by the subject, general levels of knowledge and analytic ability or clear thinking. |
|---|---|---|
| Distinction | 75-84 | Usually awarded to students whose performance goes well beyond the minimum requirements set for tasks required in assessment, and who perform well in most of the above areas. |
| Credit | 65-74 | Usually awarded to students whose performance is considered to go beyond the minimum requirements for work set for assessment. Assessable work is typically characterised by a strong performance in some of the capacities listed above. |
| Pass | 50-64 | Usually awarded to students whose performance meets the requirements set for work provided for assessment. |
| Fail | 0-49 | Usually awarded to students whose performance is not considered to meet the minimum requirements set for particular tasks. The fail grade may be a result of insufficient preparation, of inattention to assignment guidelines or lack of academic ability. A frequent cause of failure is lack of attention to subject or assignment guidelines. |
Quality assurance
For the purposes of quality assurance, Bond University conducts an evaluation process to measure and document student assessment as evidence of the extent to which program and subject learning outcomes are achieved. Some examples of student work will be retained for potential research and quality auditing purposes only. Any student work used will be treated confidentially and no student grades will be affected.
Study Information
Submission procedures
Students must check the iLearn@Bond subject site for detailed assessment information and submission procedures.
Policy on late submission and extensions
A late penalty will be applied to all overdue assessment tasks unless an extension is granted by the Lead Educator. The standard penalty will be 10% of marks awarded to that assessment per day late with no assessment to be accepted seven days after the due date. Where a student is granted an extension, the penalty of 10% per day late starts from the new due date.
Academic Integrity
Bond University‘s Student Code of Conduct Policy , Student Charter, Academic Integrity Policy and our Graduate Attributes guide expectations regarding student behaviour, their rights and responsibilities. Information on these topics can be found on our Academic Integrity webpage recognising that academic integrity involves demonstrating the principles of integrity (honesty, fairness, trust, professionalism, courage, responsibility, and respect) in words and actions across all aspects of academic endeavour.
Staff are required to report suspected misconduct. This includes all types of plagiarism, cheating, collusion, fabrication or falsification of data/content or other misconduct relating to assessment such as the falsification of medical certificates for assessment extensions. The longer term personal, social and financial consequences of misconduct can be severe, so please ask for help if you are unsure.
If your work is subject to an inquiry, you will be given an opportunity to respond and appropriate support will be provided. Academic work under inquiry will not be marked until the process has concluded. Penalties for misconduct include a warning, reduced grade, a requirement to repeat the assessment, suspension or expulsion from the University.
Feedback on assessment
Feedback on assessment will be provided to students according to the requirements of the Assessment Procedure Schedule A - Assessment Communication Procedure.
Whilst in most cases feedback should be provided within two weeks of the assessment submission due date, the Procedure should be checked if the assessment is linked to others or if the subject is a non-standard (e.g., intensive) subject.
Accessibility and Inclusion Support
Support is available to students where a physical, mental or neurological condition exists that would impact the student’s capacity to complete studies, exams or assessment tasks. For effective support, special requirement needs should be arranged with the University in advance of or at the start of each semester, or, for acute conditions, as soon as practicable after the condition arises. Reasonable adjustments are not guaranteed where applications are submitted late in the semester (for example, when lodged just prior to critical assessment and examination dates).
As outlined in the Accessibility and Inclusion Policy, to qualify for support, students must meet certain criteria. Students are also required to meet with the Accessibility and Inclusion Advisor who will ensure that reasonable adjustments are afforded to qualifying students.
For more information and to apply online, visit BondAbility.
Additional subject information
As part of the requirements for Business School quality accreditation, the Bond Business School employs an evaluation process to measure and document student assessment as evidence of the extent to which program and subject learning outcomes are achieved. Some examples of student work will be retained for potential research and quality auditing purposes only. Any student work used will be treated confidentially and no student grades will be affected.
Subject curriculum
-
Introduction to Neural Networks with Multi-Layer Perceptrons (MLPs)
The ecosystem around the use of neural networks is discussed and the foundational architecture of Multi-Layer Perceptron (MLP) models is introduced with emphasis on forward and backward passes through a network. The Keras framework is used to build our first models of the course.
-
Working with MLPs
The inter-connected elements associated with practically developing neural network models are introduced and explored. These include the requisite format of data, important processing steps, the use of mini batches for training, data splitting, and the role of hyperparameter tuning.
-
Hyperparameters, Tuning, and Models for a Purpose
The idea of regularisation, specific regularisation methods and their hyperparameters are introduced and discussed in the context of the bias-variance trade-off. Key methods for automating the tuning of the many hyperparameters associated with MLPs are explored and contrasted, including grid search, random search, and Bayesian optimisation. The importance of considering the context of the model development process is highlighted with respect to the alignment between the underlying problem, data, and type of model employed.
-
Introduction to Convolutional Neural Networks (CNNs)
The architecture of convolutional neural networks (CNNs) are motivated by the characteristics of data in computer vision problems and the core elements of CNNs are introduced. This includes the combination of convolution operations, pooling layers, and fully-connected layers that make up the CNNs used in practice. The size of datasets associated with computer vision problems is discussed with respect to the need for data generators in model training.
-
Visual Explainability and Improving CNNs
Visual methods for explaining model predictions and for model diagnostics with CNNs are introduced. This includes filter visualisations, activation maximisation, saliency maps, and class activation maps. Several innovations for training highly performant and deep CNNs are discussed, including data augmentation, batch normalisation, and residual connections.
-
CNN Case Studies and Transfer Learning
Major CNNs in the history of deep learning are described with particular emphasis on how certain design decisions have evolved over time and how the pursuit of deeper models has driven success. The topic of transfer learning, in which the features learned by a model trained in one context can be portable to other contexts, is introduced. Feature extraction and fine-tuning methodologies are contrasted.
-
Sequential Data and Recurrent Neural Networks (RNNs)
The characteristics of sequential data in many problems are described and the challenges these characteristics pose for MLPs are demonstrated. Desirable characteristics of models for sequential data motivate the discussion of recurrent neural networks. Discussion of the forward and backward passes for these models illustrates the practical challenges with training such models.
-
Improving RNNs
Several innovations for improving RNNs are discussed. These include variations on the RNN architecture that promote long-term memory, namely Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, and bidirectional layers. This also includes discussion of how dropout and skip connections can be applied effectively in the context of RNNs.
-
Embeddings and Natural Language Processing
Natural language data is extremely common, and often serves as the canonical example of problems involving sequential data. The challenges posed by natural language for computational modelling and the use of embeddings for dense representations are discussed. Four workflows for working with textual data and embeddings are introduced, in which pretrained embeddings or problem-specific embeddings are used. The Word2Vec algorithm for learning embeddings is outlined.
-
Attention in RNNs
The idea of the attention mechanism to allow a network to selectively attend to important information in an input sequence is introduced in the context of RNNs. The details of the attention mechanism are first introduced in encoder-decoder frameworks, before they are generalised to the broader ideas of using key, query, and value vectors with scoring, weighting, and summarising steps. Keras’ utility for implementing attention mechanisms on custom problems and visualising attention weights is explored.
-
Transformers
The components of the influential transformer architecture proposed in 2017 is introduced. In discussing the components, important innovations such as self-attention, layer normalisation for sequential data, positional encodings, and multi-headed attention are described.