Ivan Aerlic

Ivan Aerlic

Carlton, Victoria, Australia
2K followers 500+ connections

Join to view profile

About

Senior Salesforce Developer at Coles Group, PD II and Application Architect certified…

Activity

Join now to see all activity

Experience

  • Coles Group Graphic

    Coles Group

    Australia

  • -

    United States

  • -

    Brisbane, Queensland, Australia

  • -

    Melbourne, Victoria, Australia

  • -

    Melbourne, Victoria, Australia

Education

  • La Trobe University Graphic

    La Trobe University

    -

    Activities and Societies: N/A

    Throughout my time studying The Bachelor of Information Technology at Latrobe University, I made sure to take subjects that would give me a good foundation in programming. My academic record is dotted with a diverse theme of topics that start at web development and C++ and go on to Java and advanced principles of business analysis, where I learned about the process that goes into creating a user story. I was also taught how to construct basic and intermediate level networks that used industrial…

    Throughout my time studying The Bachelor of Information Technology at Latrobe University, I made sure to take subjects that would give me a good foundation in programming. My academic record is dotted with a diverse theme of topics that start at web development and C++ and go on to Java and advanced principles of business analysis, where I learned about the process that goes into creating a user story. I was also taught how to construct basic and intermediate level networks that used industrial grade Cisco routers, and performed the setup of hardware and software within a wide range of network typologies.

    WAM : 83.13

Licenses & Certifications

Publications

Projects

  • G-Universal-CLIP

    -

    4th place solution for the Google Universal Image Embedding Kaggle Challenge. Instance-Level Recognition workshop at ECCV 2022

Honors & Awards

  • CommonLit - Evaluate Student Summaries

    Kaggle

    Position : 2/2060
    Medal : Gold Medal
    Category : Natural Language Processing

    This competition was organized to discover how deep learning AI models could be used to grade student summary text.

    The competition lasted a total of three months with 2000+ contestants.

    The task was solved by implementing an ensemble of DeBerta transformers using the Huggingface library. Some of the techniques that were used include : layer-wise learning rate decay, Long short-term memory…

    Position : 2/2060
    Medal : Gold Medal
    Category : Natural Language Processing

    This competition was organized to discover how deep learning AI models could be used to grade student summary text.

    The competition lasted a total of three months with 2000+ contestants.

    The task was solved by implementing an ensemble of DeBerta transformers using the Huggingface library. Some of the techniques that were used include : layer-wise learning rate decay, Long short-term memory models, data augmentation (using LLMs) and pseudo-labeling.

    This competition helped me better understand how transformer can be used to solve NLP tasks.

  • BirdCLEF 2023

    Kaggle

    Position : 9/1189
    Medal : Gold Medal
    Category : Computer Vision

    This was a 3 month competition focusing on audio engineering and audio classification. But it was completed as a computer vision task.

    Our team implemented a solution using state of the art CNN models. We put together a script using the TIMM repository. This allowed us to train the EfficientNet and Convnext networks using transfer learning.

    Some of the techniques we used include : pseudo-labeling, model…

    Position : 9/1189
    Medal : Gold Medal
    Category : Computer Vision

    This was a 3 month competition focusing on audio engineering and audio classification. But it was completed as a computer vision task.

    Our team implemented a solution using state of the art CNN models. We put together a script using the TIMM repository. This allowed us to train the EfficientNet and Convnext networks using transfer learning.

    Some of the techniques we used include : pseudo-labeling, model distillation and data augmentation. All of these processes were expertly tuned to secure a 9th place finish.

  • Kaggle : RSNA Screening Mammography Breast

    Kaggle

    Position : 19/1687
    Medal : Silver Medal
    Category : Computer Vision

    Worked on a team of three Data Scientists. The task was to predict breast
    cancer cases among Mammography images.
    We used the Timm repository of models to train an ensemble of CNN models.
    Convnext, EfficientNet B3, and Eca_nfnet_l0 were among the models used
    on the list.
    We pre-trained the models on alternative Mammography datasets. We
    used heavy augmentation to regularize the outputs of the CNN…

    Position : 19/1687
    Medal : Silver Medal
    Category : Computer Vision

    Worked on a team of three Data Scientists. The task was to predict breast
    cancer cases among Mammography images.
    We used the Timm repository of models to train an ensemble of CNN models.
    Convnext, EfficientNet B3, and Eca_nfnet_l0 were among the models used
    on the list.
    We pre-trained the models on alternative Mammography datasets. We
    used heavy augmentation to regularize the outputs of the CNN models.
    Mixup and Mosaic were used. We also implement ROI cropping with YOLO
    v5 on the training data.
    Tricks like label smoothing and AUXILARY classes were used.

  • Kaggle : Google Universal Image Embedding

    Kaggle

    Position : 4/1022
    Medal : Gold Medal
    Category : Computer Vision

    Worked on a global competition for Google. The competition lasted 3
    months, contained over 1000+ teams and had a prize pool of $50,000.

    We worked with CLIP Transformer models. Created image embeddings.
    Performed dimensionality reduction using PCA. Spent dozens of hours
    preprocessing data, choosing dataset compositions and planning data
    validation strategies.

  • Kaggle : Predicting Effective Arguments

    Kaggle

    Position : 7/1557
    Medal : Gold Medal
    Category : NLP

    I was on a team of 5 Data Scientists from all over the world. We worked on a
    task of predicting the effectiveness of arguments in student written essays.
    We used the Huggingface library and DeBerta based models for this
    competition. Layer-wise learning rate decay was used. We made use of
    Stochastic Weighted Average as a way of regularizing the weights. We used
    pre-training techniques such as MLM and random word…

    Position : 7/1557
    Medal : Gold Medal
    Category : NLP

    I was on a team of 5 Data Scientists from all over the world. We worked on a
    task of predicting the effectiveness of arguments in student written essays.
    We used the Huggingface library and DeBerta based models for this
    competition. Layer-wise learning rate decay was used. We made use of
    Stochastic Weighted Average as a way of regularizing the weights. We used
    pre-training techniques such as MLM and random word insertion. We reused
    models that had been used in past competitions that were trained on similar
    datasets as a form of transfer learning. And finally we used Adversarial
    Weight Propagation to make the transformer rigorous to small changes in
    the input text.

  • Kaggle : CommonLit Readability Prize

    Kaggle

    Position : 15/3,633
    Medal : Gold Medal
    Category : NLP

    For the CommonLit Readability Prize competition I worked on a team of 5
    Data Scientists. The competition was about scoring the difficulty of a reading
    text for grades 3-12.
    For this task we used the Huggingface library. We used several transformers :
    Roberta, DeBerta, Electra and Funnel Transformer. The models were
    fine-tuned using layer-wise learning rate decay.

  • Kaggle : Shopee - Price Match Guarantee

    Kaggle

    Position : 17/2,426
    Medal : Silver Medal
    Category : Computer Vision & NLP

    I worked as a single Data Scientist on this competition which lasted about 3
    months. The goal was to match duplicate products by comparing the
    image and text.
    For this task Huggingface and the Timm repository were used to create
    image and text embeddings. Vit-Transformer and EfficienNet CNN
    architectures were combined to create image
    embeddings. For the text, distilled versions of Roberta…

    Position : 17/2,426
    Medal : Silver Medal
    Category : Computer Vision & NLP

    I worked as a single Data Scientist on this competition which lasted about 3
    months. The goal was to match duplicate products by comparing the
    image and text.
    For this task Huggingface and the Timm repository were used to create
    image and text embeddings. Vit-Transformer and EfficienNet CNN
    architectures were combined to create image
    embeddings. For the text, distilled versions of Roberta and Bert were used to
    create embedding vectors.
    This competition required cosine similarity search to be computed across
    thousands of rows. I used the FAISS library from Facebook to rapidly
    compare vectors and measure their distance.

Recommendations received

2 people have recommended Ivan

Join now to view

More activity by Ivan

View Ivan’s full profile

  • See who you know in common
  • Get introduced
  • Contact Ivan directly
Join to view full profile

Other similar profiles

Explore more posts

Explore collaborative articles

We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.

Explore More

Others named Ivan Aerlic

1 other named Ivan Aerlic is on LinkedIn

See others named Ivan Aerlic

Add new skills with these courses

See all courses